Limits may change without notice, including due to capacity constraints. When there’s a large increase in activity in Gemini Apps, we may change limits to maintain a high standard of quality. [1]
Such a beautiful paragraph to read, a company that prioritizes their current customers and focus on keeping them satisfied instead of just focusing on fast growth.
It thought for about 12 minutes and then told me I had exhausted my daily quota. (The next day Fable did the same task in 3m)
If you want to buy a plan for K3 do NOT buy the 20$ one.
I suspect that the reason it has so many parameters is the same reason that compute optimal xLSTMs have some many parameters, and the success of this model makes me a bit unhappy that we haven't gotten an xLSTM-style model of huge size developed in Europe.
Obviously these guys are very pragmatic, they're probably not committed to anything other than what works on their internal evaluations, so they still have ordinary attention layers in the model and so on, and one can't be guaranteed that the people who come up with a good model then do the engineering in an ideal way, but I still think the success of Kimi shows what could have been if we had enough big supercomputers for LLM training and made them available to the right people-- because this is basically Hochreiter's thing. It's RNNs, or well, mostly RNNs.
I just do the OpenRouter thing. My use of LLMs is narrow enough that cost is a negligible consideration either way.
Is there a good source for total token consumption between the various labs / openrouter / etc?
Back in January with K2.5 it felt like I could use it a lot more.
emphasis on _new_ subscriptions but otherwise yes that’s exactly it.
I always hated how the "Login" button is smaller on every website than the "Sign Up" button.
According to their web interface, I'm also on 23% of my weekly usage. That feels crazy, as I do a lot more than that on the 20 USD Cursor plan, and never got even a warning.
I thought OpenCode was doing something wrong, but you had the same issue with Kimi Code.
there's always a push to make the models smarter but everyone was (and still is) convinced that if we just throw more hardware at the problem and then just tell the ai to re-read its entire ''thought'' process several dozen times it'll somehow all work out and... yeah.
and it kinda works but they'll guzzle your funds very quickly now.
unless serious strides are made in efficency this problem's only gonna get worse and worse.
mLSTM layers are completely parallel (the update rule for the cell state is C_t = f_t C_{t-1} + i_t x_t where f_t and i_t are gates that can be computed from x_t alone, x_t is the input at time t and this means that you can compute F_s = \prod_{s<t} f_s as fast as a cumsum and an exponential and then compute C_t = \sum_{s<t} F_s i_t x_t, again as a cumsum, so it's as good as if though it were parallel). I think the sLSTM layers that are the other component of the xLSTM have something else like this and presumably there's some trick also to training the Kimi "delta attention" RNN.
I'm not sure whether this is hardware or optimization dependent to some degree, but I get the impression that good custom kernels are an important part of this kind of thing.
I always assumed that was because it's the least common flow. Once you've signed up, you're logged in, so the only time you should ever need the "login" button is on a new computer, or after logging out to switch accounts, or if the service offers a "remember me" checkbox (don't do that) and doesn't check it by default (definitely don't do that), or if the user genuinely wants to not be remembered (e.g. they used private browsing).
If you're having to use "login" on any regular basis, something in the overall flow needs improving.
There is a small link below the tiles for these plans saying something "looking to use it for code? Use Kimi code" and upon clicking you'll get a basically identically looking subscription page with different sub prices and more quota
Non coding plans:
* Plus (19 "base agent" -- wahtever that might be)
* Pro (39 x2 )
* Max (99 x6)
* Ultra (199 x12)
Coding plans:
* Starter (19 - "base coding")
* Explorer (49 - x4)
* Expert (99 x10)
* Master (219 x25)
I stumbled on that when i tried Kimi K2.5 back when that was released.
It's also worth mentioning that the $20 get you a 0.25 priority for API usage - because the others get x4 priority. So get ready for a _slow_ experience
I doubt my workflows match the profile of "daily driver" as you're conceiving of it. I'm mostly in the IDE chat box asking for a hand here and there, but no "long horizon" / big vision / architecture things. Occasionally, I'll solicit a general opinion, quite abstractly, which isn't very token-intensive. My contexts are tiny and I aim to keep them that way.
The only token-intensive thing I do is occasional analyses of open-source codebases along the lines of "what is the actual behaviour when message X comes in and is malformed like this?", or something else similarly far-reaching. For that, I use DeepSeek's Reasonix, which they savvily dropped at the exact moment the sting of Claude API pricing became a sensation. It's a very cache-oriented harness, and delivers on its promise of being very economical. I doubt I've spent more than $10 with DeepSeek over the months I've used it. When I want a harness to inhale a codebase and give me an answer, Reasonix has been the go-to for a while now, and it's great. DeepSeek-V4-Flash is enough for many simple things, though I just run it on DeepSeek-V4-Pro because it's so cheap that I don't care.
The plan I got is called "Moderato", which somehow isn't any of those 8... I'll try to read more about it.
I couldn't find any info on how they (siliconflow) quantized the model. I think some of the other ineference service aggregators might offer it as well, already.
I hope they wire up new hardware quickly to handle the demand.
I don't think it would even be that expensive. I think with EU chips we could do it for 20 billion EUR + 5 billion EUR per year. After all, an H200 doesn't cost 40k to manufacture, probably one-tenth, and if we are making the chips we can get the money circulating just as the Americans have.
Add in some state effort to create suitable training data-- maybe encourage academics to create some things that are AI-friendly, and let all the people trying have that. There would hardly be anyone in the game other than us if we went for this package.
We already have VSORA. Maybe Euclyd is ready soon too. If enough are ordered we have 85% of the inference need gone, so that whatever accelerators are freed can be repurposed for supercomputing and training? Maybe we could have some organized effort to make many of the VSORA and Euclyd chips so that the chips we need for training are freed up while we wait for our own training chips.
Edit: So in the end I think the thing we need to do is to order enough VSORA Jotunn8 cards and enough Euclyd cards that we flood the inference market so much that the price of H200eds etc. drops and we can buy them second hand and assemble huge supercomputers out of them. This may require some diplomatic effort from the commission, something like 20B EUR almost immediately and some coordination, but it's easy, it's doable and it requires only being a little bit decisive. Since these cards are faster less memory will be tied up in serving the models, so it'd have a positive effect on memory prices too, which in turn would have a positive effect on European supercomputer centers, who are currently of course constrained by high memory prices.
Obviously, not relevant if your reason for being here is to ferret out the excitement about K3 specifically. But if this is your first time trying Kimi, try K2.6 or K2.7! If my experience is any indication, you'll be blown away by those, too.
It’s starting to sound like Claude.
Edit: I checked and it is similar to linear delta net uses by qeen.
I remember a few high profile failures; but perhaps on average these things work out?
So often it'll cut off just as it's about to finish.
The token limits do feel a bit less than I get with Anthropic Max 5x, but maybe that's because I've mostly been running it on Max reasoning (oh and Anthropic is also temporarily boosting the limits, who knows, it's hard to keep track of all of this stuff exactly) and there's plenty of tasks where High is still close enough in performance. The token limits still feel a bit more generous proportionally to the price compared to what I got when trying out the 65 USD tier of GLM Coding Subscription with GLM 5.2, and that was with the ZCode usage discount as well, though I did enjoy that harness.
Plus, if I decide to go with Kimi's annual pricing, then it'd come out to only around 159 USD per month or 139 EUR per month, which is really good and pretty close to what I pay Anthropic anyways: https://www.kimi.com/help/membership/membership-pricing
That’s still more frequent per customer than signing up.
It only makes sense when prioritizing growth over your long-term customers.
Total usage
12.85%
Resets in 2026-08-17
5-hour usage
71.97%
Resets in 07-20 03:14
7-day usage
31.29%
Resets in 07-24 10:14
> Codex, translate the entire PostgreSQL codebase to idiomatic Rust, make no mistakes.
That said, yes, it is pretty common for me to wait 10, 20 and sometimes even 30 minutes without steering the model or looking at what its doing. I usually write a pretty detailed prompt at the start that describes the issue, the usecases, the testing to do and the definition of success; then I sometimes ask it to draft a plan and give it a read, but after that I just press enter, let it run and come back when it's done.
In the end I do a manual review, both of the code and the functionality, but most of the time I get exactly what I wanted.
But if I were to look in on them they’d be giving feedback during that time, but they wouldn’t need steering.
If my AI prompt takes more than 5 minutes to execute, I'm almost certain to stop it and break up the task or be more specific about what files it should look at or what subtask it should tackle first. The people who let it run for hours are absolutely not getting anything worth keeping in the end.
Moderato: 19 USD or 15 USD annually
Allegretto: 39 USD or 31 USD annually
Allegro: 99 USD or 79 USD annually
Vivace: 199 USD or 159 USD annually
However, they're saying that the plans might get split up again:
> New Membership Plans Coming Soon
> Kimi and Kimi Code benefits will be separated. Existing subscribers are unaffected.
> You can still buy the current plan before launch.
Same as with the Tweet:
> Going forward, we'll also split membership into two more focused plans: Kimi Membership for Kimi Web, App, and Work; and Kimi Code Membership for coding workflows. This will help us match compute more precisely and keep the experience stable.
Problem is I don't know whether to pull the trigger on the annual Vivace now or wait for after the split - whether it's gonna be a cash grab afterwards on the account of the demand, or whether I might save some money if I buy it after the split.
The day after that, Zig broke compatibility, fix that.
I've found 6 minutes or so the sweet spot for upper bound with 5.6 Sol.
And it sounds like the OPs query above requires scanning throughout a large portion of the codebase, which will inherently consume a large number of input tokens. No locality to it.
Today we needed to migrate to a cheaper CI/CD provider and do a better robotic for a particular task. GPT-5.6 Sol.
This was definitely truer with older models but isn't necessarily the case now.
They frequently do other things apart from navel-gazing that take a lot of time but get good results, like spinning up subagents to solve some hairy task in a loop.
Stopping/distrusting long-running AIs is a habit I've had to unlearn myself.
I frequently get good results from a 30+ minute Fable session, when I've asked it to do something complex (e.g. run the QA tester in a loop and eliminate all crashes, one commit per crash fixed)
The main failure point for me with long running Fable sessions now is just that it might hit a safety guardrail and downgrade to Opus midway.
This varies based on the maturity of my projects but most of my sessions live in VMs with very specific egress rules and permissions.
It works how _i_ want, which is really refreshing.