I find it really surprising that you’re fine with low end models for coding - I went through a lot of open-weights models, local and "local", and I consistently found the results underwhelming. The glm-4.7 was the smallest model I found to be somewhat reliable, but that’s a sizable 350b and stretches the definition of local-as-in-at-home.

How are you using the huge models locally?

Why don't you try it out in Opencode? It's possible to hook up the openrouter api, and some providers have started to host it there [1]. It's not yet available in opencode's model list [2].

Opencode's /connect command has a big list of providers, openrouter is on there.

[1] https://openrouter.ai/qwen/qwen3-coder-next

[2] https://opencode.ai/docs/zen/#endpoints

I configured Claude Code to use a local model (ollama run glm-4.7-flash) that runs really well on a 32G M2Pro macmini. Maybe my standards are too low, but I was using that combination to clean up the code, make improvements, and add docs and tests to a bunch of old git repo experiment projects.

It feels more like Haiku level than Sonnet 4.5 from my playing with it.

what's the token per seconds speed?

They run fairly well for me on my 128GB Framework Desktop.

We need a new word, not "local model" but "my own computers model" CapEx based

This distinction is important because some "we support local model" tools have things like ollama orchestration or use the llama.cpp libraries to connect to models on the same physical machine.

That's not my definition of local. Mine is "local network". so call it the "LAN model" until we come up with something better. "Self-host" exists but this usually means more "open-weights" as opposed to clamping the performance of the model.

It should be defined as ~sub-$10k, using Steve Jobs megapenny unit.

Essentially classify things as how many megapennies of spend a machine is that won't OOM on it.

That's what I mean when I say local: running inference for 'free' somewhere on hardware I control that's at most single digit thousands of dollars. And if I was feeling fancy, could potentially fine-tune on the days scale.

A modern 5090 build-out with a threadripper, nvme, 256GB RAM, this will run you about 10k +/- 1k. The MLX route is about $6000 out the door after tax (m3-ultra 60 core with 256GB).

Lastly it's not just "number of parameters". Not all 32B Q4_K_M models load at the same rate or use the same amount of memory. The internal architecture matters and the active parameter count + quantization is becoming a poorer approximation given the SOTA innovations.

What might be needed is some standardized eval benchmark against standardized hardware classes with basic real world tasks like toolcalling, code generation, and document procesing. There's plenty of "good enough" models out there for a large category of every day tasks, now I want to find out what runs the best

Take a gen6 thinkpad P14s/macbook pro and a 5090/mac studio, run the benchmark and then we can say something like "time-to-first-token/token-per-second/memory-used/total-time-of-test" and rate this as independent from how accurate the model was.

I run Qwen3-Coder-30B-A3B-Instruct gguf on a VM with 13gb RAM and a 6gb RTX 2060 mobile GPU passed through to it with ik_llama, and I would describe it as usable, at least. It's running on an old (5 years, maybe more) Razer Blade laptop that has a broken display and 16gb RAM.

I use opencode and have done a few toy projects and little changes in small repositories and can get pretty speedy and stable experience up to a 64k context.

It would probably fall apart if I wanted to use it on larger projects, but I've often set tasks running on it, stepped away for an hour, and had a solution when I return. It's definitely useful for smaller project, scaffolding, basic bug fixes, extra UI tweaks etc.

I don't think "usable" a binary thing though. I know you write lot about this, but it'd be interesting to understand what you're asking the local models to do, and what is it about what they do that you consider unusable on a relative monster of a laptop?

SWE-Bench Pro consists of 1865 tasks. https://arxiv.org/abs/2509.16941 Qwen3-Coder-Next solved 44.3% (826 or 827) of these tasks. To solve a single task, it took between ≈50 and ≈280 agent turns, ≈150 on average. In other words, a single pass through the dataset took ≈280000 agent turns. Kimi-K2.5 solved ≈84 fewer tasks, but also only took about a third as many agent turns.

You're replying to a bot, fyi :)

I wonder if the future in ~5 years is almost all local models? High-end computers and GPUs can already do it for decent models, but not sota models. 5 years is enough time to ramp up memory production, consumers to level-up their hardware, and models to optimize down to lower-end hardware while still being really good.

I'm thinking the next step would be to include this as a 'junior dev' and let Opus farm simple stuff out to it. It could be local, but also if it's on cerebras, it could be realllly fast.

I have the same experience with local models. I really want to use them, but right now, they're not on par with propietary models on capabilities nor speed (at least if you're using a Mac).

I can't get Codex CLI or Claude Code to use small local models and to use tools. This is because those tools use XML and the small local models have JSON tool use baked into them. No amount of prompting can fix it.

In a day or two I'll release my answer to this problem. But, I'm curious, have you had a different experience where tool use works in one of these CLIs with a small local model?

> I still haven't experienced a local model that fits on my 64GB MacBook Pro and can run a coding agent like Codex CLI or Claude code well enough to be useful

I've had mild success with GPT-OSS-120b (MXFP4, ends up taking ~66GB of VRAM for me with llama.cpp) and Codex.

I'm wondering if maybe one could crowdsource chat logs for GPT-OSS-120b running with Codex, then seed another post-training run to fine-tune the 20b variant with the good runs from 120b, if that'd make a big difference. Both models with the reasoning_effort set to high are actually quite good compared to other downloadable models, although the 120b is just about out of reach for 64GB so getting the 20b better for specific use cases seems like it'd be useful.

I experimented with the Q2 and Q4 quants. First impression is that it's amazing we can run this locally, but it's definitely not at Sonnet 4.5 level at all.

Even for my usual toy coding problems it would get simple things wrong and require some poking to get to it.

A few times it got stuck in thinking loops and I had to cancel prompts.

This was using the recommended settings from the unsloth repository. It's always possible that there are some bugs in early implementations that need to be fixed later, but so far I don't see any reason to believe this is actually a Sonnet 4.5 level model.

If it sounds too good to be true…

How's the Strix Halo? I'd really like to get a local inference machine so that I don't have to use quantized versions of local models.

what are your impressions?

I'm getting similar numbers on NVIDIA Spark around 25-30 tokens/sec output, 251 token/sec prompt processing... but I'm running with the Q4_K_XL quant. I'll try the Q8 next, but that would leave less room for context.

I tried FP8 in vLLM and it used 110GB and then my machine started to swap when I hit it with a query. Only room for 16k context.

I suspect there will be some optimizations over the next few weeks that will pick up the performance on these type of machines.

I have it writing some Rust code and it's definitely slower than using a hosted model but it's actually seeming pretty competent. These are the first results I've had on a locally hosted model that I could see myself actually using, though only once the speed picks up a bit.

I suspect the API providers will offer this model for nice and cheap, too.

> Then Anthropic started saying I’m not allowed to use my Claude Code subscription with my preferred tools

To be clear, since this confuses a lot of people in every thread: Anthropic will let you use their API with any coding tools you want. You just have to go through the public API and pay the same rate as everyone else. They have not "blocked" or "banned" any coding tools from using their API, even though a lot of the clickbait headlines have tried to insinuate as much.

Anthropic never sold subscription plans as being usable with anything other than their own tools. They were specifically offered as a way to use their own apps for a flat monthly fee.

They obviously set the limits and pricing according to typical use patterns of these tools, because the typical users aren't maxing out their credits in every usage window.

Some of the open source tools reverse engineered the protocol (which wasn't hard) and people started using the plans with other tools. This situation went on for a while without enforcement until it got too big to ignore, and they began protecting the private endpoints explicitly.

The subscription plans were never sold as a way to use the API with other programs, but I think they let it slide for a while because it was only a small number of people doing it. Once the tools started getting more popular they started closing loopholes to use the private API with other tools, which shouldn't really come as a surprise.

Easy to use a local proxy to use other models with CC. Wrote a basic working one using Claude. LiteLLM is also good. But I agree, fuck their mindset

What setup comes close to Claude Code? I am willing to rent cloude GPUs.

That’s a perfectly fine usage of content (primary substance offered by a “website”)

Access is one of my concerns with coding agents - on the one hand I think they make coding much more accessible to people who aren't developers - on the other hand this access is managed by commercial entities and can be suspended for any reason.

I can also imagine a dysfunctional future where a developers spend half their time convincing their AI agents that the software they're writing is actually aligned with the model's set of values

Still hoping IQuest-Coder gets the same treatment :)

I've had usable results with qwen3:30b, for what I was doing. There's definitely a knack to breaking the problem down enough for it.

What's interesting to me about this model is how good it allegedly is with no thinking mode. That's my main complaint about qwen3:30b, how verbose its reasoning is. For the size it's astonishing otherwise.

Ah, a spread of the individual tests makes plenty of sense! Many thanks (same goes to the other comments).

If this is genuinely better than K2.5 even at a third the speed then my openrouter credits are going to go unused.

I'm hoping so. What's amazing is that with local models you don't suffer from what I call "usage anxiety" where I find myself saving my Claude usage for hypothetical more important things that may come up, or constantly adjusting prompts and doing some manual work myself to spare token usage.

Having this power locally means you can play around and experiment more without worries, it sounds like a wonderful future.

A lot of manufacturers are bailing on consumer lines to focus on enterprise from what I've read. Not great.

Local models on your laptop will never be as powerful as the ones that take up a rack of datacenter equipment. But there is still a surprising amount of overlap if you are willing to understand and accept the limitations.

Even without leveling up hardware, 5 years is a loooong time to squeeze the juice out of lower-end model capability. Although in this specific niche we do seem to be leaning on Qwen a lot.

I do wonder if they locked things down due to people abusing their CC token.

Claude Opus 4.5 by far is the most capable development model. I've been using it mainly via Claude Code, and with Cursor.

I agree anticompetitive behavior is bad, but the productivity gains to be had by using Anthropic models and tools are undeniable.

Eventually the open tools and models will catch up, so I'm all for using them locally as well, especially if sensitive data or IP is involved.

> "content"

Sorry, but we're talking about models as content now? There's almost always a better word than "content" if you're describing something that's in tech or online.

im downloading it as we speek to try to run it on a 32gb 5090 + 128gb ddr5 i will compare it to glm 4.7-flash that was my local model of choice

You gave up some convenience to avoid voting for a bad practice with your wallet. I admire this, try to consistently do this when reasonably feasible.

Problem is, most people don't do this, choosing convenience at any given moment without thinking about longer-term impact. This hurts us collectively by letting governments/companies, etc tighten their grip over time. This comes from my lived experience.

How did you do it so fast?

Great work as always btw!

Time will tell. All this stuff will get more adoption when Anthropic, Google and OpenAI raise prices.

Hi Daniel, I've been using some of your models on my Framework Desktop at home. Thanks for all that you do.

Asking from a place of pure ignorance here, because I don't see the answer on HF or in your docs: Why would I (or anyone) want to run this instead of Qwen3's own GGUFs?

Nice! Getting ~39 tok/s @ ~60% GPU util. (~170W out of 303W per nvtop).

System info:

    $ ./llama-server --version
    ggml_vulkan: Found 1 Vulkan devices:
    ggml_vulkan: 0 = Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
    version: 7897 (3dd95914d)
    built with GNU 11.4.0 for Linux x86_64

llama.cpp command-line:

    $ ./llama-server --host 0.0.0.0 --port 2000 --no-warmup \
    -hf unsloth/Qwen3-Coder-Next-GGUF:UD-Q4_K_XL \
    --jinja --temp 1.0 --top-p 0.95 --min-p 0.01 --top-k 40 --fit on \
    --ctx-size 32768

I tried the FP8 in vLLM on my Spark and although it fit in memory, I started swapping once I actually tried to run any queries, and, yeah, could not have a context larger than 8k.

I figured out later this is because vLLM apparently de-quantizes to BF16 at runtime, so pointless to run the FP8?

I get about 30-35 tok/second using llama.cpp and a 4-bit quant. And a 200+k context, using only 50GB of RAM.

What is the difference between the UD and non-UD files?

It's getting a lot easier to do this using sub-agents with tools in Claude. I have a fleet of Mastra agents (TypeScript). I use those agents inside my project as CLI tools to do repetitive tasks that gobble tokens such as scanning code, web search, library search, and even SourceGraph traversal.

Overall, it's allowed me to maintain more consistent workflows as I'm less dependent on Opus. Now that Mastra has introduced the concept of Workspaces, which allow for more agentic development, this approach has become even more powerful.

Good results with your Q8_0 version on 96GB RTX 6000 Blackwell. It one-shotted the Flappy Bird game and also wrote a good Wordle clone in four shots, all at over 60 tps. Thanks!

Is your Q8_0 file the same as the one hosted directly on the Qwen GGUF page?

Anthropic banned my account when I whipped up a solution to control Claude Code running on my Mac from my phone when I'm out and about. No commercial angle, just a tool I made for myself since they wouldn't ship this feature (and still haven't). I wasn't their biggest fanboy to begin with, but it gave me the kick in the butt needed to go and explore alternatives until local models get good enough that I don't need to use hosted models altogether.

what do you run this on if I may ask? lmstudio, ollama, lama? which cli?

You can run plenty of models on a $10K machine or even a lot less than that, it all depends how much you want to wait for results. Streaming weights from SSD storage using mmap() is already a reality when running the largest and sparsest models. You can save even more on memory by limiting KV caching at the cost of extra compute, and there may be ways to push RAM savings even higher simply by tweaking the extent to which model activations are recomputed as needed.

For context on what cloud API costs look like when running coding agents:

With Claude Sonnet at $3/$15 per 1M tokens, a typical agent loop with ~2K input tokens and ~500 output per call, 5 LLM calls per task, and 20% retry overhead (common with tool use): you're looking at roughly $0.05-0.10 per agent task.

At 1K tasks/day that's ~$1.5K-3K/month in API spend.

The retry overhead is where the real costs hide. Most cost comparisons assume perfect execution, but tool-calling agents fail parsing, need validation retries, etc. I've seen retry rates push effective costs 40-60% above baseline projections.

Local models trading 50x slower inference for $0 marginal cost start looking very attractive for high-volume, latency-tolerant workloads.

I don't even need "open weights" to run on hardware I own.

I am fine renting an H100 (or whatever), as long as I theoretically have access to and own everything running.

I do not want my career to become dependent upon Anthropic.

Honestly, the best thing for "open" might be for us to build open pipes and services and models where we can rent cloud. Large models will outpace small models: LLMs, video models, "world" models, etc.

I'd even be fine time-sharing a running instance of a large model in a large cloud. As long as all the constituent pieces are open where I could (in theory) distill it, run it myself, spin up my own copy, etc.

I do not deny that big models are superior. But I worry about the power the large hyperscalers are getting while we focus on small "open" models that really can't match the big ones.

We should focus on competing with large models, not artisanal homebrew stuff that is irrelevant.

I won't need a heater with that running in my room.

OOM is a pretty terrible benchmark too, though. You can build a DDR4 machine that "technically" loads 256gb models for maybe $1000 used, but then you've got to account for the compute aspect and that's constrained by a number of different variables. A super-sparse model might run great on that DDR4 machine, whereas a 32b model would cause it to chug.

There's just not a good way to visualize the compute needed, with all the nuance that exists. I think that trying to create these abstractions are what leads to people impulse buying resource-constrained hardware and getting frustrated. The autoscalers have a huge advantage in this field that homelabbers will never be able to match.

Honestly I've been completely spoiled by Claude Code and Codex CLI against hosted models.

I'm hoping for an experience where I can tell my computer to do a thing - write a code, check for logged errors, find something in a bunch of files - and I get an answer a few moments later.

Setting a task and then coming back to see if it worked an hour later is too much friction for me!

If it weren't for the single em-dash (really an en-dash, used as if it were an em-dash), how am I supposed to know that?

And at the end of the day, does it matter?

"Is they key unlock here"

Opensource or local models will always heavily lag frontier.

Who pays for a free model? GPU training isn't free!

I remember early on people saying 100B+ models will run on your phone like nowish. They were completely wrong and I don't think it's going to ever really change.

People always will want the fastest, best, easiest setup method.

"Good enough" massively changes when your marketing team is managing k8s clusters with frontier systems in the near future.

Plus a long queue of yet-undiscovered architectural improvements

Cerebras already has GLM 4.7 in the code plans

I'm using this model right now in claude code with LM Studio perfectly, on a macbook pro

This model is exactly what you’d want for your resources. GPU for prompt processing, ram for model weights and context length, and it being MoE makes it fairly zippy. Q4 is decent; Q5-6 is even better, assuming you can spare the resources. Going past q6 goes into heavily diminishing resources.

Tune it out, come back in 6 months, the world is not going to end. In 6 months, you’re going to change your API endpoint and/or your subscription and then spend a day or two adjusting. Off to the races you go.

I have a 3090 and a 4090 and it all fits in in VRAM with Q4_0 and quantized KV, 96k ctx. 1400 pp, 80 tps.

Relatively, it's not that hard. There's like 4-5 "real" AI labs, who altogether manage to announce maybe 3 products max, per-month.

Compared to RISC core designs or IC optimization, the pace of AI innovation is slow and easy to follow.

This is going to be a crazy month because the Chinese labs are all trying to get their releases out prior to their holidays (Lunar New Year / Spring Festival).

So we've seen a series of big ones already -- GLM 4.7 Flash, Kimi 2.5, StepFun 3.5, and now this. Still to come is likely a new DeepSeek model, which could be exciting.

And then I expect the Big3, OpenAI/Google/Anthropic will try to clog the airspace at the same time, to get in front of the potential competition.

I'd be happy with something that's close or same as opus 4.5 that I can run locally, at reasonable (same) speed as claude cli, and at reasonable budget (within $10-30k).

Just code it yourself, you might surprise yourself :)

Pretty much every lab you can think of has something scheduled for february. Gonna be a wild one

Well there are rumors sonnet 5 is coming today, so...

Yes, the general purpose version is already supported and should have the same identical architecture

Try KimiK2.5 and DeepSeekv3.2-Speciale

1 6000 should be fine, Q6_K_XL gguf will be almost on par with the raw weights and should let you have 128k-256k context.

Yes, exactly. The discourse has been so far off the rails now.

The question I pose is this: if they're willing to start building walls this early in the game while they've still got plenty of viable competitors, and are at most 6 months ahead, how will they treat us if they achieve market dominance?

Some people think LLMs are the final frontier. If we just give in and let Anthropic dictate the terms to us we're going to experience unprecedented enshittification. The software freedom fight is more important than ever. My machine is sovereign; Anthropic provides the API, everything I do on my machine is my concern.

from what i remember, i couldnt actually use claude code with the subscription when i subscribed. i could only use it with third party tools.

eventually they added subscription support and that worked better than cline or kilo, but im still not clear what anthropic tools the subscription was actually useful for

I don't get why so much mental gymnastics is done to avoid the fact that locking their lower prices to effectively subsidize their shitty product is the anti competitive behavior.

They simply don't want to compete, they want to force the majority of people that can't spend a lot on tokens to use their inferior product.

Why build a better product if you control the cost?

This model does not fit in 12G of VRAM - even the smallest quant is unlikely to fit. However, portions can be offloaded to regular RAM / CPU with a performance hit.

I would recommend trying llama.cpp's llama-server with models of increasing size until you hit the best quality / speed tradeoff with your hardware that you're willing to accept.

The Unsloth guides are a great place to start: https://unsloth.ai/docs/models/qwen3-coder-next#llama.cpp-tu...

Not applicable... the models just process whatever context you provide to them, context management happens outside of the model and depends on your inference tool/coding agent.

No, at Q2 you are looking at a size of about 26gb-30gb. Q3 exceeds it, you might run it, but the result might vary. Best to run a smaller model like qwen3-32b/30b at Q6

> But my experience is they’re not really even close to the closed paid models.

They are usually as good as the flagship model for 12-18 months ago. Which may sound like a massive difference, because somehow it is, but it's also fairly reasonable, you don't need to live to the bleeding edge.

By the time that happens, Opus 5 and GPT-5.5 will be out. At that point will a GPT-5.2 tier open-weights model feel "good enough"? Based on my experience with frontier models, once you get a taste of the latest and greatest it's very hard to go back to a less capable model, even if that less capable model would have been SOTA 9 months ago.

Essentially the more turns you have the more the agent is likely to fail since the error compounds per turn. Agentic model are tuned for “long horizon tasks” ie being able to go many many turns on the same problem without failing.

Who cares? If you don't like it, you can fine tune.

The anticompetitive part is setting a much lower price for typical usage of Claude Code vs. typical usage of another CLI dev tool.

> Anthropic will let you use their API with any coding tools you want

No, in 2026, even with their API plan the create key is disabled for most orgs, you basically have to ask your admin to give you a key to use something other than Claude Code. You can imagine how that would be a problem.

Nah, their "moat" is CC, they are afraid that as other folks build effective coding agent, they are are going lose market share.

Taking umbrage as if it matters how I use the compute I'm paying for via the harness they want me to use it within as long as I'm just doing personal tasks I want to do for myself, not trying to power an apps API with it seems such a waste of their time to be focusing on and only causes brand perception damage with their customers.

Could have just turned a blind eye.

I wasn’t only referring to their new model, I meant their blogpost and the research behind their progress, its always a joyride to read.

I didn’t know it was this serious with the vocabulary, I’ll be more cautious in the future.

Not everyone on hn is a native english speaker...

The loss of access shows the kind of power they'll have in the future. It's just a taste of what's to come.

If a company is going to automate our jobs, we shouldn't be giving them money and data to do so. They're using us to put ourselves out of work, and they're not giving us the keys.

I'm fine with non-local, open weights models. Not everything has to run on a local GPU, but it has to be something we can own.

I'd like a large, non-local Qwen3-Coder that I can launch in a RunPod or similar instance. I think on-demand non-local cloud compute can serve as a middle ground.

It feels very close to a trade-off point.

I agree with all posts in the chain: Opus is good, Anthropic have burned good will, I would like to use other models...but Opus is too good.

What I find most frustrating is that I am not sure if it is even actual model quality that is the blocker with other models. Gemini just goes off the rails sometimes with strange bugs like writing random text continuously and burning output tokens, Grok seems to have system prompts that result in odd behaviour...no bugs just doing weird things, Gemini Flash models seem to output massive quantities of text for no reason...it is often feels like very stupid things.

Also, there are huge issues with adopting some of these open models in terms of IP. Third parties are running these models and you are just sending them all your code...with a code of conduct promise from OpenRouter?

I also don't think there needs to be a huge improvement in models. Opus feels somewhat close to the reasonable limit: useful, still outputs nonsense, misses things sometimes...there are open models that can reach the same 95th percentile but the median is just the model outputting complete nonsense and trying to wipe your file system.

The day for open models will come but it still feels so close and so far.

Likewise curious to hear how it goes! 80B seems too big for a 5090, I'd be surprised if it runs well un-quantized.

Super cool! Also with `--fit on` you don't need `--ctx-size 32768` technically anymore - llama-server will auto determine the max context size!

Interested to hear how this goes!

Society is lacking people that stand up for something. My efforts to consume less is seen as being cheap by my family, which I find so sad. I much prefer donating my money than exchanging superfluous gifts on Christmas.

Thanks! Oh Qwen3's own GGUFs also works, but ours are dynamically quantized and calibrated with a reasonably large diverse dataset, whilst Qwen's ones are not - see https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs

They can only raise prices as long as people buy their subscriptions / pay for their api. The Chinese labs are closing in on the SOTA models (I would say they are already there) and offer insane cheap prices for their subscriptions. Vote with your wallet.

Nice! Yes Q8_0 is similar - the others are different since they use a calibration dataset.

Running llama.cpp rather than vLLM, it's happy enough to run the FP8 variant with 200k+ context using about 90GB vram

Thanks! :) We're early access partners with them!

Can't speak for parent, but I've had decent luck with llama.cpp on my triple Ryzen AI Pro 9700 XTs.

At this point isn’t the marginal cost based on power consumption? At 30c/kWh and with a beefy desktop pc pulling up to half a kW, that’s 15c/hr. For true zero marginal cost, maybe get solar panels. :P

This looks like it’ll run easily on a Strix Halo (180W TDP), and be a little sluggish on previous gen AMDs (80W TDP).

I can’t be bothered to check TDPs on 64GB macbooks, but none of these devices really count as space heaters.

Yeah, that hits different.

Some people reply for their own happiness, some reply to communicate with another person. The AI won't remember or care about the reply.

I think we'll eventually find a way to make the cycle smaller, so instead of writing a stackoverflow post in 2024 and using a model trained on it in 2025 I'll be contributing to the expertise of a distributed-model-ish-thing on Monday and benefitting from that contribution on Tuesday.

When that happens, the most powerful AI will be whichever has the most virtuous cycles going with as wide a set of active users as possible. Free will be hard to compete with because raising the price will exclude the users that make it work.

Until then though, I think you're right that open will lag.

I don't know about frontier, I code nowadays a lot using Opus 4.5, in a way that I instruct it to do something (like complex refactor etc) - I like that it's really good at actually doing what its told and only occasionally do I have to fight it when it goes off the rails. It also does not hallucinate all that much in my experience (Im writing Js, YMMV with other languages), and is good at spotting dumb mistakes.

That said, I'm not sure if this capability is only achievable in huge frontier models, I would be perfectly content using a model that can do this (acting as a force multiplier), and not much else.

> People always will want the fastest, best, easiest setup method

When there are no other downsides, sure. But when the frontier companies start tightening the thumbscrews, price will influence what people consider good enough.

The calculation will probably get better for locally hosted models once investor generosity runs out for the remotely hosted models.

You mean Qwen3-Coder-Next? I haven't tried that model itself, yet, because I assume it's too big for me. I have a modest 16GB MacBook Air so I'm restricted to really small stuff. I'm thinking about buying a machine with a GPU to run some of these.

Anywayz, maybe I should try some other models. The ones that haven't worked for tool calling, for me are:

Llama3.1

Llama3.2

Qwen2.5-coder

Qwen3-coder

All these in 7b, 8b, or sometimes 30b (painfully) models.

I should also note that I'm typically using Ollama. Maybe LM Studio or llama.cpp somehow improve on this?

Are you running 120B agentic? I tried using it in a few different setups and it failed hard in every one. It would just give up after a second or two every time.

I wonder if it has to do with the message format, since it should be able to do tool use afaict.

You are describing distillation, there are better ways to do it, and it was done in the past, Deepseek distilled onto Qwen.

There have been advances recently (last year) in scaling deep rl by a significant amount, their announcement is in line with a timeline of running enough experiments to figure out how to leverage that in post training.

Importantly, this isn’t just throwing more data at the problem in an unstructured way, afaik companies are getting as many got histories as they can and doing something along the lines of, get an llm to checkpoint pull requests, features etc and convert those into plausible input prompts, then run deep rl with something which passes the acceptance criteria / tests as the reward signal.

I would not go below q8 if comparing to sonnet.

Wonder where it falls on the Sonnet 3.7/4.0/4.5 continuum.

3.7 was not all that great. 4 was decent for specific things, especially self contained stuff like tests, but couldn't do a good job with more complex work. 4.5 is now excellent at many things.

If it's around the perf of 3.7, that's interesting but not amazing. If it's around 4, that's useful.

Prompt preprocessing is slow, the rest is pretty great.

Works great for these type of MOE models. The ability to have large amounts of VRAM let you run different models in parallel easily, or to have actually useful context sizes. Dense models can get sluggish though. AMD's ROCm support has been a little rough for Stable Diffusion stuff (memory issues leading to application stability problems) but it's worked well with LLMs, as does Vulkan.

I wish AMD would get around to adding NPU support in Linux for it though, it has more potential that could be unlocked.

And it's worth pointing out that Claude Code now dispatches "subagents" from Opus->Sonnet and Opus->Haiku ... all the time, depending on the problem.

Running this thing locally on my Spark with 4-bit quant I'm getting 30-35 tokens/sec in opencode but it doesn't feel any "stupider" than Haiku, that's for sure. Haiku can be dumb as a post. This thing is smarter than that.

It feels somewhere around Sonnet 4 level, and I am finding it genuinely useful at 4-bit even. Though I have paid subscriptions elsewhere, so I doubt I'll actually use it much.

I could see configuration OpenCode somehow to use paid Kimi 2.5 or Gemini for the planning/analysis & compaction, and this for the task execution. It seems entirely competent.

Thank you for your advice have a good evening

When Alibaba succeeds at producing a GPT-5.2-equivalent model, they won't be releasing the weights. They'll only offer API access, like for the previous models in the Qwen Max series.

Don't forget that they want to make money in the end. They release small models for free because the publicity is worth more than they could charge for them, but they won't just give away models that are good enough that people would pay significant amounts of money to use them.

I buy the theory that Claude Code is engineered to use things like token caching efficiently, and their Claude Max plans were designed with those optimizations in mind.

If people start using the Claude Max plans with other agent harnesses that don't use the same kinds of optimizations the economics may no longer have worked out.

(But I also buy that they're going for horizontal control of the stack here and banning other agent harnesses was a competitive move to support that.)

In what way would it be abused? The usage limits apply all the same, they aren't client side, and hitting that limit is within the terms of the agreement with Anthropic.

> Claude Opus 4.5 by far is the most capable development model.

At the moment I have a personal Claude Max subscription and ChatGPT Enterprise for Codex at work. Using both, I feel pretty definitively that gpt-5.2-codex is strictly superior to Opus 4.5. When I use Opus 4.5 I’m still constantly dealing with it cutting corners, misinterpreting my intentions and stopping when it isn’t actually done. When I switched to Codex for work a few months ago all of those problems went away.

I got the personal subscription this month to try out Gas Town and see how Opus 4.5 does on various tasks, and there are definitely features of CC that I miss with Codex CLI (I can’t believe they still don’t have hooks), but I’ve cancelled the subscription and won’t renew it at the end of this month unless they drop a model that really brings them up to where gpt-5.2-codex is at.

How do I "abuse" a token? I pass it to their API, the request executes, a response is returned, I get billed for it. That should be the end of the conversation.

(Edit due to rate-limiting: I see, thanks -- I wasn't aware there was more than one token type.)

I'd encourage you to try the -codex family with the highest reasoning.

I can't comment on Opus in CC because I've never bit the bullet and paid the subscription, but I have worked my way up to the $200/month Cursor subscription and the 5.2 codex models blow Opus out of the water in my experience (obviously very subjective).

I arrived at making plans with Opus and then implementing with the OpenAI model. The speed of Opus is much better for planning.

I'm willing to believe that CC/Opus is truly the overall best; I'm only commenting because you mentioned Cursor, where I'm fairly confident it's not. I'm basing my judgement on "how frequently does it do what I want the first time".

As I get older I more and more view convenience as the enemy of good. Luckily (or unluckily for some) a lot of the tradeoffs we are asked to make in the name of convenience are increasingly absurd. I have an easier and easier time going without these Faustian bargains.

What am I missing here? I thought this model needs 46GB of unified memory for 4-bit quant. Radeon RX 7900 XTX has 24GB of memory right? Hoping to get some insight, thanks in advance!

UD stands for "Unsloth-Dynamic" which upcasts important layers to higher bits. Non UD is just standard llama.cpp quants. Both still use our calibration dataset.

Are you just exposing mastra cli commands to Claude Code in md context? I’d love you to elaborate on this if you have time.

There is weaponized malaise employed by these frontier model providers and I feel like that dark-pattern, what you pointed out, and others are employed to rate-limit certain subscriptions.

How did this work? The ban, I mean. Did you just wake up to find out an email and that your creds no longer worked? Were you doing things to sub-process out to the Claude Code CLI or something else?

They did ship that feature, it's called "&" / teleport from web. They also have an iOS app.

I control it with ssh and sometimes tmux (but termux+wireguard lead to a surprisingly generally stable connection). Why did you need more than that?

Yeah there's a lot of people that advocate for really slow inference on cheap infra. That's something else that should be expressed in this fidelity

Because honestly I don't care about 0.2 tps for my use cases although I've spoken with many who are fine with numbers like that.

At least the people I've talked to they talk about how if they have a very high confidence score that the model will succeed they don't mind the wait.

Essentially a task failure is 1 in 10, I want to monitor and retry.

If it's 1 in 1000, then I can walk away.

The reality is most people don't have a bearing on what this order of magnitude actually is for a given task. So unless you have high confidence in your confidence score, slow is useless

But sometimes you do...

> I do not want my career to become dependent upon Anthropic

As someone who switches between Anthropic and ChatGPT depending on the month and has dabbled with other providers and some local LLMs, I think this fear is unfounded.

It's really easy to switch between models. The different models have some differences that you notice over time but the techniques you learn in one place aren't going to lock you into a provider anywhere.

> time-to-first-token/token-per-second/memory-used/total-time-of-test

Would it not help with the DDR4 example though if we had more "real world" tests?

I don't think this is as true as you think.

People do not care about the fastest and best past a point.

Let's use transportation as an analogy. If all you have is a horse, a car is a massive improvement. And when cars were just invented, a car with a 40mph top speed was a massive improvement over one with a 20mph top speed and everyone swapped.

While cars with 200mph top speeds exist, most people don't buy them. We all collectively decided that for most of us, most of the time, a top speed of 110-120 was plenty, and that envelope stopped being pushed for consumer vehicles.

If what currently takes Claude Opus 10 minutes to do can be done is 30ms, then making something that can do it in 20ms isn't going to be enough to get everyone to pay a bunch of extra money for.

Companies will buy the cheapest thing that meets their needs. SOTA models right now are much better than the previous generation but we have been seeing diminishing returns in the jump sizes with each of the last couple generations. If the gap between current and last gen shrinks enough, then people won't pay extra for current gen if they don't need it. Just like right now you might use Sonnet or Haiku if you don't think you need Opus.

Gpt3.5 as used in the first commercially available chat gpt is believed to be hundreds of billions of parameters. There are now models I can run on my phone that feel like they have similar levels of capability.

Phones are never going to run the largest models locally because they just don't have the size, but we're seeing improvements in capability at small sizes over time that mean that you can run a model on your phone now that would have required hundreds of billions of parameters less than 6 years ago.

Yep. But this is like 10x faster; 3B active parameters.

I'm suprised there isn't more "hope" in this area. Even things like the GPT Pro models; surely that sort of reasoning/synthesis will eventually make its way into local models. And that's something that's already been discovered.

Just the other day I was reading a paper about ANNs whose connections aren't strictly feedforward but, rather, circular connections proliferate. It increases expressiveness at the (huge) cost of eliminating the current gradient descent algorithms. As compute gets cheaper and cheaper, these things will become feasible (greater expressiveness, after all, equates to greater intelligence).

> I experimented with the Q2 and Q4 quants.

Of course you get degraded performance with this.

Should be possible with optimised models, just drop all "generic" stuff and focus on coding performance.

There's no reason for a coding model to contain all of ao3 and wikipedia =)

Surely the answer is a very small proxy server between the two?

It literally always is. HN Thought DeepSeek and every version of Kimi would finally dethrone the bigger models from Anthropic, OpenAI, and Google. They're literally always wrong and average knowledge of LLMs here is shockingly low.

I got Codex CLI running against it and was sadly very unimpressed - it got stuck in a loop running "ls" for some reason when I asked it to create a new file.

llama.cpp is giving me ~35tok/sec with the unsloth quants (UD-Q4_K_XL, elsewhere in this thread) on my Spark. FWIW my understanding and experience is that llama.cpp seems to give slight better performance for "single user" workloads, but I'm not sure why.

I'm asking it to do some analysis/explain some Rust code in a rather large open source project and it's working nicely. I agree this is a model I could possibly, maybe use locally...

I’ve a 128GB m3 max MacBook Pro. Running the gpt oss model on it via lmstudio once the context gets large enough the fans spin to 100 and it’s unbearable.

It's interesting how people can be so into LLMs but dont, at the end of the day, understand they're just passing "well formatted" text to a text processor and everything else is build around encoding/decoding it into familiar or novel interfaces & the rest.

The instability of the tooling outside of the LLM is what keeps me from building anything on the cloud, because you're attaching your knowledge and work flow to a tool that can both change dramatically based on context, cache, and model changes and can arbitrarily raise prices as "adaptable whales" push the cost up.

Its akin to learning everything about beanie babies in the early 1990's and right when you think you understand the value proposition, suddenly they're all worthless.

Thanks for the pointers!

one more thing, that guide says:

> You can choose UD-Q4_K_XL or other quantized versions.

I see eight different 4-bit quants (I assume that is the size I want?).. how to pick which one to use?

    IQ4_XS
    Q4_K_S
    Q4_1
    IQ4_NL
    MXFP4_MOE
    Q4_0
    Q4_K_M
    Q4_K_XL

I think it depends on what you use it for. Coding, where time is money? You probably want the Good Shit, but also want decent open weights models to keep prices sane rather than sama’s 20k/month nonsense. Something like a basic sentiment analysis? You can get good results out of a 30b MoE that runs at good pace on a midrange laptop. Researching things online with many sources and decent results I’d expect to be doable locally by the end of 2026 if you have 128GB ram, although it’ll take a while to resolve.

If an open weights model is released that’s as capable at coding as Opus 4.5, then there’s very little reason not to offload the actual writing of code to open weight subagents running locally and stick strictly to planning with Opus 5. Could get you masses more usage out of your plan (or cut down on API costs).

I'm going in the opposite direction: with each new model, the more I try to optimize my existing workflows by breaking the tasks down so that I can delegate tasks to the less powerful models and only rely on the newer ones if the results are not acceptable.

Anticompetitive with themselves? It’s not like Claude / Anthropic have any kind of monopoly, and services companies are allowed to charge different rates for different kind of access to said service?

That’s not an Anthropic problem, that’s a problem with whomever you work for.

It should just burn quota faster then. Instead of blocking they should just mention that if you use other tools then your quota may reduce at 3x speed compared to cc. People would switch.

I would be surprised if the primary reason for banning third party clients isn't because they are collecting training data via telemetry and analytics in CC. I know CC needlessly connects to google infrastructure, I assume for analytics.

When I last checked a few months ago, Anthropic was the only provider that didn't have automatic prompt caching. You had to do it manually (and you could only set checkpoints a few times per context?), and most 3rd party stuff does not.

They seem to have started rejecting 3rd party usage of the sub a few weeks ago, before Claw blew up.

By the way, does anyone know about the Agents SDK? Apparently you can use it with an auth token, is anyone doing that? Or is it likely to get your account in trouble as well?

If that was the real reason, why wouldn't they just make it so that if you don't correctly use caching you use up more of your limit?

Thanks, I'll try those out. I've used Codex CLI itself on a few small projects as well, and fired it up on a feature branch where I had it implement the same feature that Claude Code did (they didn't see each other's implementations). For that specific case, the implementation Codex produced was simpler, and better for the immediate requirements. However, Claude's more abstracted solution may have held up better to changing requirements. Codex feels more reserved than Claude Code, which can be good or bad depending on the task.

IMHO The question is: who is in control? The user, or the profit-seeking company/control-seeking government? There is nothing we can do to prevent companies from seeking profit. What we can do is to prefer tools that we control, if that choice is not available, then tools that we can abandon when we want, over tools that remove our control AND abandoning them would be prohibitively difficult.

MoEs can be efficiently split between dense weights (attention/KV/etc) and sparse (MoE) weights. By running the dense weights on the GPU and offloading the sparse weights to slower CPU RAM, you can still get surprisingly decent performance out of a lot of MoEs.

Not as good as running the entire thing on the GPU, of course.

Seconded!

It feels like the gap between open weight and closed weight models is closing though.

The anticompetitive move would be not running their software if ‘which codex’ evaluated to showing a binary and then not allow you to use it due to its presence. Companies are allowed to set pricing and not let you borrow the jet to fly to a not approved destination. This distortion is just wrong as a premise. They are being competitive by making a superior tool and their business model is “no one else sells Claude” and they are pretty right to do this IMO.

Absolutely. I installed clawdbot for just long enough to send a single message, and it burned through almost a quarter of my session allowance. That was enough for me. Meanwhile I can use CC comfortably for a few hours and I've only hit my token limit a few times.

I've had a similar experience with opencode, but I find that works better with my local models anyway.

The subscription services have assumptions baked in about the usage patterns; they're oversubscribed and subsidized. If 100% of subscriber customers use 100% of their tokens 100% of the time, their business model breaks. That's what wholesale / API tokens are for.

> hitting that limit is within the terms of the agreement with Anthropic

It's not, because the agreement says you can only use CC.

I have literally the opposite experience and so does most of AI pilled twitter and the AI research community of top conferences (NeurIPS, ICLR, ICML, AAAI) Why does this FUD keep appearing on this site?

Edit: It's very true that the big 4 labs silently mess with their models and any action of that nature is extremely user hostile.

You can buy this product, right here: https://platform.claude.com/docs/en/about-claude/pricing

That's not the product you buy when you a Claude Code token, though.

I've tried nearly all the models, they all work best if and only if you will never handle the code ever again. They suck if you have a solution and want them to implement that solution.

I've tried explaining the implementation word and word and it still prefers to create a whole new implementation reimplementing some parts instead of just doing what I tell it to. The only time it works is if I actually give it the code but at that point there's no reason to use it.

There's nothing wrong with this approach if it actually had guarantees, but current models are an extremely bad fit for it.

What is your definition of "important" in this context?

Please consider authoring a single, straightforward introductory-level page somewhere that explains what all the filename components mean, and who should use which variants.

The green/yellow/red indicators for different levels of hardware support are really helpful, but far from enough IMO.

They have two products:

* Subscription plans, which are (probably) subsidized and definitely oversubscribed (ie, 100% of subscribers could not use 100% of their tokens 100% of the time).

* Wholesale tokens, which are (probably) profitable.

If you try to use one product as the other product, it breaks their assumptions and business model.

I don't really see how this is weaponized malaise; capacity planning and some form of over-subscription is a widely accepted thing in every industry and product in the universe?

I left a sibling comment detailing the technical side of things. I used the `Bun.spawn` API with the `terminal` key to give CC a PTY and mirrored it to my phone with xterm.js. I used SSE to stream CC data to xterm.js and a regular request to send commands out from my phone. In my mind, this is no different than using CC via SSH from my phone - I was still bound by the same limits and wasn't trying to bypass them, Anthropic is entitled to their different opinion of course.

And yeah, I got three (for some reason) emails titled "Your account has been suspended" whose content said "An internal investigation of suspicious signals associated with your account indicates a violation of our Usage Policy. As a result, we have revoked your access to Claude.". There is a link to a Google Form which I filled out, but I don't expect to hear back.

I did nothing even remotely suspicious with my Anthropic subscription so I am reasonably sure this mirroring is what got me banned.

Edit: BTW I have since iterated on doing the same mirroring using OpenCode with Codex, then Codex with Codex and now Pi with GPT-5.2 (non-Codex) and OpenAI hasn't banned me yet and I don't think they will as they decided to explicitly support using your subscription with third party coding agents following Anthropic's crackdown on OpenCode.

That's non-local. I am not interested in coding assistants that work on cloud based work-spaces. That's what motivated me to developed this feature for myself.

I used to say that Sonnet 4.5 was all I would ever need, but now I exclusively use Opus...

> Based on my experience with frontier models, once you get a taste of the latest and greatest it's very hard to go back to a less capable model, even if that less capable model would have been SOTA 9 months ago.

That's the tyranny of comfort. Same for high end car, living in a big place, etc.

There's a good work around though: just don't try the luxury in the first place so you can stay happy with the 9 months delay.

I think a lot of people care. Most decidedely not you.

Obviously. That's why I led with that statement.

Those are the quant thresholds where people with mid-high end hardware can run this locally at reasonable speed, though.

In my experience Q2 is flakey, but Q4 isn't dramatically worse.

There is: It works (even if we can't explain why right now).

If we knew how to create a SOTA coding model by just putting coding stuff in there, that is how we would build SOTA coding models.

But... but... I need my coding model to be able to write fanfiction in the comments...

Now I wonder how strong the correlation between coding performance and ao3 knowledge is in human programmers. Maybe we are on to something here /s

That might work, but I keep seeing people talk about this, so there must be a simple solution that I'm over-looking. My solution is to write my own minimal and experimental CLI that talks JSON tools.

Yes sadly that sometimes happens - the issue is Codex CLI / Claude Code were designed for GPT / Claude models specifically, so it'll be hard for OSS models directly to utilize the full spec / tools etc, and might get loops sometimes - I would maybe try the MXFP4_MOE quant to see if it helps, and maybe try Qwen CLI (was planning to make a guide for it as well)

I guess until we see the day OSS models truly utilize Codex / CC very well, then local models will really take off

Nobody has been saying they'd be dethroned. We're saying they're often "good enough" for many use cases, and that they're doing a good job of stopping the Big Guys from creating a giant expensive moat around their businesses.

Chinese labs are acting as a disruption against Altman etcs attempt to create big tech monopolies, and that's why some of us cheer for them.

Yeah I got 35-39tok/sec for one shot prompts, but for real-world longer context interactions through opencode it seems to be averaging out to 20-30tok/sec. I tried both MXFP4 and Q4_K_XL, no big difference, unfortunately.

--no-mmap --fa on options seemed to help, but not dramatically.

As with everything Spark, memory bandwidth is the limitation.

I'd like to be impressed with 30tok/sec but it's sort of a "leave it overnight and come back to the results" kind of experience, wouldn't replace my normal agent use.

However I suspect in a few days/weeks DeepInfra.com and others will have this model (maybe Groq, too?), and will serve it faster and for fairly cheap.

Laptops are fundamentally a poor form factor for high performance computing.

That's why you can use latest open coding models locally that reportedly reached the performance of Sonet-4.5 so almost SOTA. And then you can think of tricks like I mentioned above to directly manipulate GPU RAM for context cleanup when needed which is not possible with cloud models unless their provider enables that.

The I-prefix stands for Imatrix smoothing in the quantization. It trades a little more accuracy for speed than other quant styles. The _0 and _1 quants are older, simpler quants that are very accurate but kinda slow. The K quants, in my limited understanding, primarily quantize at the specified bit depth, but will bump certain important areas higher, and less used parts lower. It generally performs better while providing similar accuracy to the _1 quants. MXFP4 is specific to Nvidia, so I can't use it on my AMD hardware. It's supposed to be very efficient. The UD part includes more of Unsloth's speed optimizations.

Also, depending on how much regular system RAM you have, you can offload mixture-of-expert models like this, keeping only the most important layers on your GPU. This may let you use larger, more accurate quants. That is functionality that is supported by llama.cpp and other frameworks and is worth looking into how to do.

Much appreciated, but I mean more around "what do the error bars in the figure represent" than what the turn scaling itself is.

I think I like coding models that know a lot about the world. They can disambiguate my requirements and build better products.

That's what Meta thought initially too, training codellama and chat llama separately, and then they realized they're idiots and that adding the other half of data vastly improves both models. As long as it's quality data, more of it doesn't do harm.

Besides, programming is far from just knowing how to autocomplete syntax, you need a model that's proficient in the fields that the automation is placed in, otherwise they'll be no help in actually automating it.

Yeah, Apple hardware don't seem ideal for LLMs that are large, give it a go with a dedicated GPU if you're inclined and you'll see a big difference :)

What does it mean for U.S. AI firms if the new equilibrium is devs running open models on local hardware?

Anticompetitive behavior has been normalized in our industry, doesn't make it not anticompetitive. It's a restriction that's meant to make it harder to compete with other parts of their offering. The non-anticompetitive approach would be to offer their subscription plans with a certain number of tokens every month, and then make Claude Code the most efficient with the tokens, to let it compete on its own merits.

If you launch enough tasks in parallel you aren't going to care that 1 in 10 failed, as long as the other 9 are good. Just rerun the failed job whenever you get around to it, the infra will still be getting plenty of utilization on the rest.

> It's really easy to switch between models. The different models have some differences that you notice over time but the techniques you learn in one place aren't going to lock you into a provider anywhere.

We have two cell phone providers. Google is removing the ability to install binaries, and the other one has never allowed freedom. All computing is taxed, defaults are set to the incumbent monopolies. Searching, even for trademarks, is a forced bidding war. Businesses have to shed customer relationships, get poached on brand relationships, and jump through hoops week after week. The FTC/DOJ do nothing, and the EU hasn't done much either.

I can't even imagine what this will be like for engineering once this becomes necessary to do our jobs. We've been spoiled by not needing many tools - other industries, like medical or industrial research, tie their employment to a physical location and set of expensive industrial tools. You lose your job, you have to physically move - possibly to another state.

What happens when Anthropic and OpenAI ban you? Or decide to only sell to industry?

This is just the start - we're going to become more dependent upon these tools to the point we're serfs. We might have two choices, and that's demonstrably (with the current incumbency) not a good world.

Computing is quickly becoming a non-local phenomenon. Google and the platforms broke the dream of the open web. We're about to witness the death of the personal computer if we don't do anything about it.

This is the assumption of a hard plateu we can effectively optimize forever towards while possible we havn't seen it.

Again my point is "good enough" changes as possibilities open. Marketing teams running entire infra stacks is an insane idea today but may not be in the future.

You could easily code with a local model similar to gpt 4 or 3 now but I will 10-100x your performance with a frontier model and that will fundamentally not change.

Hmmm but maybe there's an argument of a static task. Once a model hits that ability of that specific task you can optimize it into a smaller model. So I guess I buy the argument for people working on statically capped conplexity tasks?

PII detection for example, a <500M model will outperform a 1-8B param model on that narrow task. But at the same time just a pii detection bot is not a product anymore. So yes a opensource one does it but as a result its fundamentally less valuable and I need to build higher and larger products for the value?

Maybe, but even that fourth-order metric is missing key performance details like context length and model size/sparsity.

The bigger takeaway (IMO) is that there will never really be hardware that scales like Claude or ChatGPT does. I love local AI, but it stresses the fundamental limits of on-device compute.

Sure but the moment you can use that small model locally its capabilities are no longer differntiated or valuable no?

I supose the future will look exacrly like now. Some mixture of local and non local.

I guess my argument is that market dominated by local doesn't seem right and I think the balance will look similar to what it is right now

Mode like open local models are becoming "good enough".

I got stuff done with Sonnet 3.7 just fine, it did need a bunch of babysitting, but still it was a net positive to productivity. Now local models are at that level, closing up on the current SOTA.

When "anyone" can run an Opus 4.5 level model at home, we're going to be getting diminishing returns from closed online-only models.

I didn't like the existing SSH applications for iOS and I already have a local app that I made that I have open 24/7, so I added a screen that used xterm.js and Bun.spawn with Bun.Terminal to mirror the process running on my Mac to my phone. This let me add a few bells and whistles that a generic SSH client wouldn't have, like notifications when Claude Code was done working etc.

right, but ChatGPT might not exist at some point, and if we don't force feed the open inference ecosystem and infrastructure back into the mouths of the AI devourer that is this hype cycle, we'll simply be accepting our inevitable, painful death

The G in GPT stands for Generalized. You don't need that for specialist models, so the size can be much smaller. Even coding models are quite general as they don't focus on a language or a domain. I imagine a model specifically for something like React could be very effective with a couple of billion parameters, especially if it was a distill of a more general model.

Cerebras is already 200-800 tps, do you need even faster ?

It seems like a lot of the benefits of SOTA models are from data though, not architecture? Won't the moat of the big 3/4 players in getting data only grow as they are integrated deeper into businesses workflows?

Wow, that is very surprising and alarming. I wish Anthropic would have made a more public statement as to why they blocked other harnesses.

That's on Anthropic for selling a mirage of limits they don't want people to actually reach for.

It's within their capability to provision for higher usage by alternative clients. They just don't want to.

Probably because all of the major providers are constantly screwing around with their models, regardless of what they say.

Yes, I only plan/implement on fully AI projects where it's easy for me to tell whether or not they're doing the thing I want regardless of whether or not they've rewritten the codebase.

For actual work that I bill for, I go in with intructions to do minimal changes, and then I carefully review/edit everything.

That being said, the "toy" fully-AI projects I work with have evolved to the point where I regularly accomplish things I never (never ever) would have without the models.

> It's not, because the agreement says you can only use CC.

it's like Apple: you can use macOS only on our Macs, iOS only on iPhones, etc. but at least in the case of Apple, you pay (mostly) for the hardware while the software it comes with is "free" (as in free beer).

There are domains of programming (web front end) where lots of requests can be done pretty well even when you want them done a certain way. Not all, but enough to make it a great tool.

Claude Code supports using API credits, and you can turn on Extra Usage and use API credits automatically once your session limit is reached.

This confused me for a while, having two separate "products" which are sold differently, but can be used by the same tool.

The green/yellow/red indicators are based on what you set for your hardware on huggingface.

> Anthropic is entitled to their different opinion of course.

I'm not so sure. It doesn't sound like you were circumventing any technical measures meant to enforce the ToS which I think places them in the wrong.

Unless I'm missing some obvious context (I don't use Mac and am unfamiliar with the Bun.spawn API) I don't understand how hooking a TUI up to a PTY and piping text around is remotely suspicious or even unusual. Would they ban you for using a custom terminal emulator? What about a custom fork of tmux? The entire thing sounds absurd to me. (I mean the entire OpenCode thing also seems absurd and wrong to me but at least that one is unambiguously against the ToS.)

Oh we wrote about it here: https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs

That's a box plot, so those are not error bars but a visualization of the distribution of a metric (min, max, median, 25th percentile, 75th percentile).

The benchmark consists of a bunch of tasks. The chart shows the distribution of the number of turns taken over all those tasks.

But as far as I know, that was way before tool calling was a thing.

I'm more bullish about small and medium sized models + efficient tool calling than I'm about LLMs too large to be run at home without $20k of hardware.

The model doesn't need to have the full knowledge of everything built into it when it has the toolset to fetch, cache and read any information available.

For the tasks in SWE-Bench Pro they obtained a distribution of agent turns, summarized as the box plot. The box likely describes the inter-quartile range while the whiskers describe the some other range. You'd have to read their report to be sure. https://en.wikipedia.org/wiki/Box_plot

I used it for a few mins and it burned 7M tokens. Wish there was a way to see where it's going!

(There probably is, but I found it very hard to make sense of the UI and how everything works. Hard to change models, no chat history etc.?)

> The subscription services have assumptions baked in about the usage patterns; they're oversubscribed and subsidized.

Selling dollars for $.50 does that. It sounds like they have a business model issue to me.

I am curious to see how this will pan out long-term. Is the quality gap of Opus-4.5 over GPT-5.2 large enough to overcome the fact that OpenAI has merged these two bullet points into one? I think Anthropic might have bet on no other frontier lab daring to disconnect their subscription from their in-house coding agent and OpenAI called their bluff to get some free marketing following Anthropic's crackdown on OpenCode.

Oh good idea! In general UD-Q4_K_XL (Unsloth Dynamic 4bits Extra Large) is what I generally recommend for most hardware - MXFP4_MOE is also ok

But... Claude Code is already cloud-based. It relies on the Anthropic API. Your data is all already being ingested by them. Seems like a weird boundary to draw, trusting the company's model with your data but not their convenience web ui. Being local-only (ie OpenCode & open weights model running on your own hw) is consistent, at least.

> Anthropic is entitled to their different opinion of course.

It’d be cool if Anthropic were bound by their terms of use that you had to sign. Of course, they may well be broad enough to fire customers at will. Not that I suggest you expend any more time fighting this behemoth of a company though. Just sad that this is the state of the art.

I generally prefer a coding model that can google for the docs, but separate models for /plan and /build is also a thing.

So, if I rent out my bike to you for an hour a day for really cheap money and I do so a 50 more times to 50 others, so that my bike is oversubscribed and you and others don't get your hours, that's OK because it is just capacity planning on my side and widely accepted? Good to know.

What are some good GPUs to look for if you're getting started?

OpenAI isn’t cornering the market on DRAM for kicks…

Thats what i want and orchestrator model that operates with a small context and then very specialized small models for react etc

I'll be that guy: the "G" in GPT stands for "Generative".

That's a good point. I'm not familiar enough with the various moats to comment.

I was just talking at a high level. If transformers are HDD technology, maybe there's SSD right around the corner that's a paradigm shift for the whole industry (but for the average user just looks like better/smarter models). It's a very new field, and it's not unrealistic that major discoveries shake things up in the next decade or less.

I have a feeling the different harnesses create new context windows instead of using one. The more context windows you open up with Claude the quicker your usage goes poof.

How did they even know you did this? I cannot imagine what cause they could have for the ban. They actively want folks building tooling around and integrating with Claude Code.

See, the market is investing like _that will never happen_.

If they die there will be so much hardware released to do other tasks.

Yes! I don't try to read agent tokens as they are generated, so if code generation decreases from 1 minute to 6 seconds, I'll be delighted. I'll even accept 10s -> 1s speedups. Considering how often I've seen agents spin wheels with different approaches, faster is always better, until models can 1-shot solutions without the repeated "No, wait..." / "Actually..." thinking loops

This is how every cloud service and every internet provider works. If you want to get really edgy you could also say it's how modern banking works.

Without knowing the numbers it's hard to tell if the business model for these AI providers actually works, and I suspect it probably doesn't at the moment, but selling an oversubscribed product with baked in usage assumptions is a functional business model in a lot of spaces (for varying definitions of functional, I suppose). I'm surprised this is so surprising to people.

It will also be interesting to see which model is more sustainable once the money fire subsidy musical chairs start to shake out; it all depends on how many whales there are in both directions I think (subscription customers using more than expected vs large buys of profitable API tokens).

Let me introduce you to Citibike?

Also, this is more like "I sell a service called take a bike to the grocery store" with a clause in the contract saying "only ride the bike to the grocery store." I do this because I am assuming that most users will ride the bike to the grocery store 1 mile away a few times a week, so they will remain available, even though there is an off chance that some customers will ride laps to the store 24/7. However, I also sell a separate, more expensive service called Bikes By the Hour.

My customers suddenly start using the grocery store plan to ride to a pub 15 miles away, so I kick them off of the grocery store plan and make them buy Bikes By the Hour.

It sucks and I wish it were different, but it is not so different from trying to get support at Meta or Google. If I was an AI grifter I could probably just DM a person on Twitter and get this sorted, but as a paying customer, it's wisest to go where they actually want my money.

Yes, correct. Essentially every single industry and tool which rents out capacity of any system or service does this. Your ISP does this. The airline does this. Cruise lines. Cloud computing environments. Restaurants. Rental cars. The list is endless.

Is there some indication on how the different bit quantization affect performance? IE I have a 5090 + 96GB so I want to get the best possible model but I don't care about getting 2% better perf if I only get 5 tok/s.

It is not a moral stance. I just prefer to have my files of my personal projects in one place. Sure I sync them to GitHub for backup, but I don't use GitHub for anything else in my personal projects. I am not going to use a workflow which relies on checking out my code to some VM where I have to set everything up in a way where it has access to all the tools and dependencies that are already there on my machine. It's slower, clunkier. IMO you can't beat the convenience of working on your local files. When I used my CC mirror for the brief period where it worked, when I came back to my laptop, all my changes were just already there, no commits, no pulls, no sync, nothing.

> separate models for /plan and /build

I had not considered that, seems like a great solution for local models that may be more resource-constrained.

As others pointed out, every business that sells capacity does this, including your ISP provider.

They could, of course, price your 10GB plan under the assumption that you would max out your connection 24 hours a day.

I fail to see how this would be advantageous to the vast majority of the customers.

I have some bad news for you about your home internet connection.

I have no idea. The alternative is that my account just happened to be on the wrong side of their probably slop-coded abuse detection algorithm. Not really any better.

I'm just riding the VC powered wave of way-too-cheap online AI services and building tools and scaffolding to prepare for the eventual switch to local models =)

Perhaps not tasks you get the opportunity to do.

Your job might be assigned to some other legal entity renting some other compute.

If this goes as according to some of their plans, we might all be out of the picture one day.

If these systems are closed, you might not get the opportunity to hire them yourself to build something you have ownership in. You might be cut out.

> until models can 1-shot solutions without the repeated "No, wait..." / "Actually..." thinking loops

That would imply they'd have to be actually smarter than humans, not just faster and be able to scale infinitely. IMHO that's still very far away..

>Without knowing the numbers it's hard to tell if the business model for these AI providers actually works

It'll be interesting to see what OpenAI and Anthropic will tell us about this when they go public (seems likely late this year--along with SpaceX, possibly)

If they rely on this to be competitive, I have serious doubts they will survive much longer.

There are already many serious concerns about sharing code and information with 3rd parties, and those Chinese open models are dangerously close to destroying their entire value proposition.

Don't forget gyms and other physical-space subscriptions. It's right up there with razor-and-blades for bog standard business models. Imagine if you got a gym membership and then were surprised when they cancelled your account for reselling gym access to your friends.

The Business model is Uber. It doesn't work unless you corner the market and provide a distinct value replacement.

The problem is, there's not a clear every-man value like Uber has. The stories I see of people finding value are sparse and seem from the POV of either technosexuals or already strong developer whales leveraging the bootstrapy power .

If AI was seriously providing value, orgs like Microsoft wouldn't be pushing out versions of windows that can't restart.

It clearly is a niche product unlike Uber, but it's definitely being invested in like it is universal product.

Ah okay, that makes sense. Sorry they pulled the plug on you!

You can configure aider that way. You get three, in fact: an architect model, a code editor model, and a quick model for things like commit messages. Although I'm not sure if it's got doc searching capabilities.

> selling an oversubscribed product with baked in usage assumptions is a functional business model in a lot of spaces

Being a common business model and it being functional are two different things. I agree they are prevalent, but they are actively user hostile in nature. You are essentially saying that if people use your product at the advertised limit, then you will punish them. I get why the business does it, but it is an adversarial business model.

Well, if the service price were in any way tied to the cost of transmitting bytes, then even the 24hr scenarios would likely see a reduction in cost to customers. Instead we have overage fees and data caps to help with "network congestion", which tells us all how little they think of their customers.

It takes download time + 1 minute to test speed yourself, you can try different quants, it's hard to write down a table because it depends on your system ie. ram clock etc. if you go out of gpu.

I guess it would make sense to have something like max context size/quants that fit fully on common configs with gpus, dual gpus, unified ram on mac etc.

Testing speed is easy yes, I'm mostly wondering about the quality difference between Q6 vs Q8_K_XL for example.

I haven't done benchmarking yet (plan to do them), but it should be similar to our post on DeepSeek-V3.1 Dynamic GGUFs: https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs