I installed it and it's none of that. It is a mere wrapper around small local LLM models. And, it's not even multi-modal! Anyone could've one-shotted this in Claude in an hour (I'm not exaggerating).
What's the target audience here? Your average person doesn't care about the privacy value proposition (at least not by severely sacrificing chat model's quality). And users who do want that control can already install LMStudio/Llama.cpp (which is dead simple to setup).
The actual release product should've been what's described in "What's next" section.
> Instead of general chat, we shape Ensu to have a more specialized interface, say like a single, never-ending note you keep writing on, while the LLM offers suggestions, critiques, reminders, context, alternatives, viewpoints, quotes. A second brain, if you will.
> A more utilitarian take, say like an Android Launcher, where the LLM is an implementation detail behind an existing interaction that people are already used to.
> Your agent, running on your phone. No setup, no management, no manual backups. An LLM that grows with you, remembers you, your choices, manages your tasks, and has long-term memory and personality.
This probably could have been one-shotted with Sonnet, not even Opus. Given how over indexed they are on LLM coding, Haiku might even be able to do it.
This is actually an interesting coding model benchmark task now that I think about it.
Absolutely no one called them crazy.
They also have a TOTP auth app?
If their photos app stopped crashing and they pursued basic feature parity between their iOS and desktop apps (IMO table stakes for a photo sync service) I'd have no issue recommending them. Instead, it seems like every so often they just branch off into a new direction, leaving the existing products unfinished. It's like Mozilla-level lack of focus.
Then moved to pocket pal now for local llm.
Come onnnnnn. I would rather read a one line "Check out our offline llm" rather than a whole press release of slop.
This looks very neat. I'm not familiar with the nitty gritty of AI so I really don't understand how it can reply so quickly running on an iPhone 16. But I'm not even going to bother searching for details because I don't want to read slop.
I have a phone in a drawer I could install termux and ollama on over tailscale and then I'd have an always on llm for super light tasks.
I do really long for a private chat bot but I simply don't have access to the hardware required. Sadly I think it's going to be years to get there..
If Ente is reading this : please add requirements to make it run (how many RAM, etc.)
Have a comparison chart to Ollama, LMStudio, LocalAI, Exo, Jan.AI, GPT4ALL, PocketPal, etc.
https://github.com/Arthur-Ficial/apfel
Apple Ai on the command line
I've found https://github.com/alichherawalla/off-grid-mobile-ai but haven't tried anything in this space yet.
Going to give this a try...
Ideally if you "participate" in the network, you would get "credits" to use it proportionally to how much GPU power you have provided to the network. Or if you can't, then buy credits (payment would be distributed as credits to other participants).
That way we could build huge LLMs that area really open and are not owned by any network.
I would LOVE to participate in building that as well.
However, it’s a bit confusing because, for example, a larger LLM model was downloaded to my smartphone than to my computer. It would probably make the most sense if the app simply categorized devices into five different tiers and then, depending on which performance tier a device falls into, downloaded the appropriate model and simply informed the user of the performance tier. Over time, it would then be possible to periodically replace the LLM for each tier with better ones, or to redefine the device performance tiers based on hardware advancements.
How does it compare to Jan AI for example? or LM Studio? or ????
A little bit of cleanup on their site to break out "Ente, our original photo sharing app" from the rest of their apps would do wonders, because I had to search around on the announcement to find the download for this app, which feels about like trying to find the popular Ente Auth app on their website
It requires a Firefox add-on to act as a bridge: https://addons.mozilla.org/en-US/firefox/addon/ai-s-that-hel...
There is honestly not much to test just yet, but feel free to check it out here, provide feedback on the idea: https://codeberg.org/Helpalot/ais-that-helpalot
The essence works, I was able to let it make a simple summary on CMS content. So next is making it do something useful, and making it clear how other plugins could use it.
If it's so great, why is there so little viscera documenting it's greatness? Just lots and lots of words.
I think they did. If you start the download and then open the sidebar and/or background the app, the download progress bar disappears and is replaced by the download button. If you press the download button again, the progress bar reappears at the correct point.
I find that Claude often makes little statefulness mistakes like that. Human developers do too, but the slower and more iterative nature of human development makes it more likely that that would get caught.
We have not seen a tidal wave of untechnical people vibe coding up their own software solutions.
There is truly nothing original here and the product doesn't have a chance in hell of earning money. Local LLMs on-device will be dominated by the device vendors, whose control of the hardware stack combined with their ability to subsidize billions of dollars of machine learning research gives them an unfair advantage. Apple knows what the next generation of silicon will deliver, and their ML engineers are already hard at work building models that will be highly optimized for that silicon a year or two ahead of time. Open source models are really great and are backed by well funded labs; however, delivering these models on-device in a way that pleases users will never be easier than it is for the vendors of the devices.
Plus, device vendors have ways of making money from local LLMs that third-party app providers do not. They can make their local LLM free and earn money on the hardware play, without skipping a beat on the billions of dollars of ongoing R&D. I don't see how third party app vendors make money here when they will be competing with the decent, totally free alternative that Apple and Google (and Samsung etc.) will load on in the next year or two.
When my little brother who is a drummer, and has never even looked at "code" before, had claude on-shot a python app that let him download songs on youtube, extract the stems, collect tempo/key/etc information, then feed that into his DAW, all without ever looking at code, knowing what any of it did, etc., I knew that we were about to see a LOT of single-use applications.
I'm not against it, honestly. I have always written little one-off scripts and apps that accomplished something faster than manually, now that those one-shots are possible with an LLM in seconds sometimes, it makes all my personal scripts so much easier... that said, I definitely read the scripts that are output, and attempt to step through them in a debugger before assuming it is all good.
That to me is more valuable than code vibe coded by Claude in one afternoon.
(Though I think this announcement is sufficiently unpleasant I'm starting to reconsider)
But sure, making money with standalone "local first is our headline feature" will be incredibly hard against those, no doubt about that. In light of the limited quality of what local models can achieve, the privacy bonus just won't compel many to pay. But that only means that this "morning with Claude" you are suggesting might be just the right amount of investment for the result you'd realistically expect. But is that so bad? I'd argue the reverse: bundling up the low hanging fruit but not by some hobbyist who will lose interest two weeks on but by a company big enough the keep it going while small enough to not be a VC furnace that will inevitably turn on users once the runway runs out (*), that's an opportunity to fill a niche few others can. Valuable for users who don't want to roll their own deployment of open source models (can't, or unwilling to commit to keeping them up to date, assuming that Ente does keep that ball rolling), and also valuable for the company of the investment actually is so low that it pays by raising awareness for their other products that apparently do earn them money.
(*) I was googling around a little wondering if they actually are as close to bootstrapped as they seem on the surface, and yes, that's supposedly the core idea [0], but despite that they also took 100 kUSD in "non-diluting" (basically a gift then?) from Mozilla with the explicit goal "to promote independent AI and machine learning" [1]. So not a CEO whim but following up to a promise made earlier. If they actually did avoid spending all that money on a one-off but went smaller planning to keep it current for a longer time horizon, I'd congratulate them on an excellent choice.
[0] https://ente.com/blog/5-years-of-ente/
[1] https://ente.io/blog/mozilla-builders/
The hn discussion for [1] seems to be completely missing the point, that Mozilla program isn't about funding an image host (yeah, I'd also prefer if Mozilla focused on the Browser and perhaps Thunderbird, but the foundation is what it is): https://news.ycombinator.com/item?id=41681666
LLMs are too important to be left to big tech. There is a gap between frontier models and models that can run on your device, but local models improve each day, and once they cross a certain capability threshold, they will be good enough for most purposes; and will come with full privacy and control.
Based on these assumptions, we've been working on Ensu, Ente's offline LLM app. Today is our first release.
Download it here.
In the rest of this post, we'll explain why we think the assumptions hold, what we're doing, and how you can get involved.
LLMs are too important to be left to big tech. We've written in depth about this earlier, here and here.
Briefly, those posts come at it from two angles:
If you're someone who hates LLMs, you would still be able to recognize in clearer moments of thought that LLMs are a technology that can't just be wished away.
If you're someone who finds joy in interacting with LLMs, you would recognize the lack of privacy and the high dependency (arbitrary bans, content shaping, non-portable memory) you have on centralized providers.
And in both cases it is also clear that LLMs can be used to manipulate people en masse. Ergo, we can't be at the mercy of big tech controlling them.
The issue is that there is a capability gap between large centralized models and smaller models that can be run offline on your device.
But we're problem solvers, and this is not our first rodeo. When we first started Ente Photos, it seemed unthinkable that we'd be able to deliver face recognition, person clustering and natural language image search all running locally on your device. People called us crazy.
It took many years, but we did it. Our users enjoy these features every day. Everything is done locally on device, and also synced, end-to-end encrypted, across all your devices. Full privacy, full control, without loss of convenience; technology in service of people, not as a tool of surveillance.
In the same vein, while we have been itching for a long time to do something about local LLMs, it is only recently that smaller models are becoming feasible to run on consumer devices. We now think there are actionable steps we can take.
This is where the second assumption comes in. While smaller decentralized models improve every day, so do the larger centralized models. However, we think the gap is not what is important - instead, it is about a threshold, and about how the model's capabilities are used. Once smaller models will cross a certain threshold, they will be sufficient to provide joy, utility and convenience in the life of people.
Today we're releasing Ensu. It is a chatgpt-like app that runs completely on your device with full privacy and zero cost. Soon, you'll also be able to backup and sync your chats across your devices by connecting your Ente account (or self hosting), with full end-to-end encryption.
This is not the beginning, nor is this the end. This is just a checkpoint.
Ensu is currently an Ente Labs project. For now, we want to only iterate on the product and its direction, without bringing pricing and stability too early into the picture.
Just to set expectations right, it is currently not as powerful as ChatGPT or Claude Code. Still, it is already quite fun! Here are some things we've been doing with it:
Introspecting about thoughts we wouldn’t risk putting into a non-private LLM.
Talking about books (Ensu currently doesn't have web search, but you'll be surprised how well it knows classics like the Gita or the Bible)
Jabbering with it on flights when there is no internet.
The app is open source, and available for iOS, Android, macOS, Linux and Windows. We also have an experimental web based version. Image attachments are also supported. The core logic is written in Rust, and for each platform we have native (mobile) and Tauri (desktop) apps that use the same shared logic.
We've already implemented (optional) E2EE syncing and backups so that you can access your chats across devices. This uses the Ente account you already have, and it can also be self hosted just like Ente Photos. However, at the last minute we decided not to enable sync in the checkpoint we're releasing today. That's the story of the next section.
We're viewing Ensu as a journey. There is a precise destination - a private, personal LLM with encrypted sync - however the path to it is hazy. There are multiple directions we could take:
Instead of general chat, we shape Ensu to have a more specialized interface, say like a single, never-ending note you keep writing on, while the LLM offers suggestions, critiques, reminders, context, alternatives, viewpoints, quotes. A second brain, if you will.
A more utilitarian take, say like an Android Launcher, where the LLM is an implementation detail behind an existing interaction that people are already used to.
Your agent, running on your phone. No setup, no management, no manual backups. An LLM that grows with you, remembers you, your choices, manages your tasks, and has long-term memory and personality.
For now we will just wait a while for feedback before taking the next step. And because these future directions might change the persistence architecture, we've delayed enabling sync.
When sync does arrive, your existing local chats will get backed up and sync too.
We would love your feedback. The next steps are unclear, and we want you to influence what we build. Tell us what you want, and we'll make it. You can write to us at [email protected], or join our Discord and head over to the #ensu channel.
You can download Ensu here.
I do agree that more local LLM options are always better.
But where are they! https://ente.com/about
Small team, rooting for them
When the comments here say "there's no value because anyone could've compiled llama.cpp", you can see how detached from reality these people are.
Even jumping through the hoops to get an app on Play Store and Apple Store — an app that I can tell my friends to look up and download — is worth a lot.
An app that is also is available on Mac and PC, mind you.
I'm an ex-Google/Meta/Microsoft/Roblox software engineer, and I couldn't be bothered to do any of that.
Neither could the rest of HN. But I'm not the one complaining about lack of novelty or value in this proposition.
What I'm missing is a way to create and use Passkeys across devices. My use case does not support creating a new Passkey on every device, I need to sync them via servers I control. The system that supports that will be the system that I migrate to.
The people who got tired of waiting for "any half capable engineer" to do so.
I'd love to know a few more local LLM apps that are available on Android and iOS and Mac/PC under the same branding that I can point my non-technical friends to as a ChatGPT alternative that works offline (but still has sync across the devices).
Could you recommend a few?
EDIT: and there are long-standing bugs like this one, unaddressed: https://github.com/ente-io/ente/issues/3087
Helping non-technical people get off of ChatGPT.com and using increasingly better local models seems worth celebrating and continued iteration.
Does this seem sound?
Expressly harvesting creds through a 2FA app seems a little more direct.
Either LFM2.5-1.6B-4bit or Qwen3.5-2B-8bit or Qwen3.5-4B-4bit
This was posted the other day, but only briefly made the front page - seems kinda like what you’re talking about
https://github.com/ente-io/ente/blob/f254af939ff6950b63edf5f... Here is the system prompt, kinda embarassing
For when wordpress doesn't have enough exploits and bugs as it is. Also why bother with wordpress in the first place if you're already having an LLM spit out content for you ?
Also: "Your AI agent can now create, edit, and manage content on WordPress.com" https://wordpress.com/blog/2026/03/20/ai-agent-manage-conten...
This does the same for language models.
https://en.wikipedia.org/wiki/Comparison_of_OTP_applications
They just store tokens, without other FA at "worst" you get locked of your account but nobody else has access either. You're also supposed to, as good practice, not be limited to token generation and typically have a dozen or so of recovery tokens. Also if they were somewhat not working at doing the 1 task they should do, namely generate tokens, then you won't be able to use them so it won't even be added.
So... I might be missing something, can you please explain what worries you and why I should thus worry too?
Though, I don't see any references to Gemma at all in the open source code...
You can check the code for exploits yourself. And other than that it's just your LLM talking to your own website.
> Also why bother with wordpress in the first place
Weird question, but sure, I use WordPress, because I have a website that I want to run with a simple CMS that can also run my custom Wordpress plugins.
I'm talking about connecting Ollama to your wordpress.
Not via MCP or something that's complicated for a relatively normal user. But thanks for the link.
Although the ability to use large models "for free" sounds pretty rad.
Here’s where it was added to PrivacyGuides - https://github.com/privacyguides/privacyguides.org/issues/36.... The person opening the issue is the CEO of ente. So the CEO of ente gets his company mentioned in PrivacyGuides back when it was new and that makes it more legit?
I would really like to know what people use these small and tiny models for. If any high-karma users are reading it, would you consider posting Ask HN?
Claude Code is a Desktop app as well.
If the new Wordpress feature would allow for connecting to Ollama, then there is no need anymore for my plugin. But I don't see that in the current documentation.
So for now, I see my solution being superior for anyone who doesn't have a paid subscription, but has a decent laptop, that would like to use an LLM 'for free' (apart from power usage) with 100% privacy on their website.
very limited amount of use cases, perhaps. As a generalized chat assistant? I'm not sure you'd be able to get anything of value out from them, but happy to be proven otherwise. I have all of those locally already, without fine-tuning, what use case could I try right now where any of those are "very effective"?
So you look down you see a tortoise. It's crawling towards you.
For the user it's just important that the small grimlin that sits in the Ente app is not as smart as the grimlin that sits in the Claude app.
> Use Claude Code where you work
> Desktop Termianl IDE WEb and iOS Slack
Not that it is important any way ¯\_(ツ)_/¯
You can use a small coding model to produce working code with a deterministic workflow (ex: state machine) if you carefully prune the context and filter down what it can do per iteration. Instead of letting it "reason" through an ever growing history, you give it distinct piecemeal steps with tailored context.
I think this can be generalized to:
Anything that can be built from small, well understood pieces and can be validated and fixed step by step. Then the challenge becomes designing these workflows and automating them.
(I'm not there yet, but one thing I have in mind might be a hybrid approach where the planning is produced by a more expensive model. The output it has to produce are data driven state machines or behavior trees (so they can be validated deterministically). Then it offloads the grunt work to a small, local model. When it's done, the work gets checked etc.)