I like the approach of running everything locally. I'm strongly of the opinion that the privacy angle for local models is going to keep getting stronger and more relevant. The amount of articles that come out about accidents happening because of people handing too much context to cloud models the more self reinforcing this will become.
It's only half of the solution though. If the models are trained in a closed way, they can prioritize values encoded during training even if that's not what you want (example: ask the open Chinese models about Tiananmen). It's not beyond imagining that these models would e.g. try to send your data to authorities or advertisers when their training says so, even if you run them locally.
So the full solution would be models trained in an open verifiable way and running locally.
Another angle is when you're passing untrusted content to the AI service, e.g. anything from using it to crawl websites to spam-detection on new forum user posts.
You can trigger the the service's ToS violation or worse, get tipped off to law enforcement for something you didn't even write.
local is best for privacy, but i personally think you don't need to go local.
anthropic, google, openai etc, decided that their consumer ai plans would not be private. partly to collect training data, the other half to employ moderators to review user activity for safety.
we trust that human moderators will not review and flag our icloud docs, onedrive or gmail, or aggregate such documents into training data for llms. it became the norm that an llm is somehow not private. it became a norm that you can't opt out of training, even on paid plans (see meta and google); or if you can opt out of training, you can't opt out of moderation.
cloud models with a zero retention privacy policy are private enough for almost everyone, the subscriptions, google search, ai search engines are either 'buying' your digital life or covering themselves for legal reasons.
you can and should have private cloud services, and if legal agreement is not enough, cryptographic attestation is already used in compute, with AWS nitro enclaves and other providers.
I personally think everyone should default to using local resources. Cloud resources should only be used for expansion and be relatively bursty rather than the default.
For about two years I experimented with writing local apps using local LLMs, but I often had to blend in a commercial web search API to make my little experiments useful.
I pay $13/month for Proton’s Lumo+ private chat LLM that contains an excellent built-in web search tool. I use it for everything non-technical, even just simple searching for local businesses, etc.
As an enthusiastic reader of books like Privacy is Power and Surveillance Capitalism, it feels good to have a private tool that is ready at hand.
do you have any provider recommendations? I've experimented with this on runpod serverless, but I've been meaning to dig deeper before I feel comfortable with personal data.
I saw a service named Phala, which claims to be actually no-knowledge to server side (I think). It was significantly more expensive, but interesting to see it's out there. My thought was escaping the data-collection-hungry consumer models was a big win.
> anthropic, google, openai etc, decided that their consumer ai plans would not be private. partly to collect training data, the other half to employ moderators to review user activity for safety.
That's two halves of "why", sure.
Another interesting half would be that those companies have US military officers on their boards, and LLMs are the ultimate voluntary data collection platform, even better trojan horses than smartphones.
Yet another "half" could be how much enterprise value might be found by datamining for a minute or two... may I suggest reading a couple of Martha Wells books.
The other thing, is encrypted inferencing a thing/service currently? I want to run my own models locally just because if I'm going to be chatting to it about my day to day life why send it to a server in plaintext.
> I like the approach of running everything locally. I'm strongly of the opinion that the privacy angle for local models is going to keep getting stronger and more relevant.
I’ve seen several projects like this that offer a network server with access to these Apple models. The danger is when they expose that, even on a loop port, to every other application on your system, including the browser. Random webpages are now shipping with JavaScript that will post to that port. Same-origin restrictions will stop data flow back to the webpage, but that doesn’t stop them from issuing commands to make changes.
Some such projects use CORS to allow read back as well. I haven’t read Apfel’s code yet, but I’m registering the experiment before performing it.
They offer it as an option but default it to false! This is still a --footgun option but it’s the least unsafe version I’ve seen yet! Well done, Apfel authors.
FWIW this was the status quo (webpage could ping arbitrary ports but not read data, even with CORS protections) - but it is changing.
This is partially in response to https://localmess.github.io/ where Meta and Yandex pixel JS in websites would ping a localhost server run by their Android apps as a workaround to third-party cookie limits.
So things are getting better! But there was a scarily long time where a rogue JS script could try to blindly poke at localhost servers with crafty payloads, hoping to find a common vulnerability and gain RCE or trigger exfiltration of data via other channels. I wouldn't be surprised if this had been used in the wild.
There is a CORS preflight check for POST requests that don't use form-encoding. It would be somewhat surprising if these weren't using JSON (though it wouldn't be that surprising if they were parsing submitted JSON instead of actually checking the MIME-type which would probably be bad anwyay)
Isn't there a CORS preflight check for this? In most cases. I guess you could fashion an OG form to post form fields. But openai is probably a JSON body only.
The default scenario should be secure. If the local site sends permissive CORS headers bets may be off. I would need to check but https->http may be a blocker too even in that case. Unless the attack site is http.
With the Claude bug, or so it is known, burning through tokens at record speed, I gave alternative models a try and they're mostly ... interchangeable. I don't know how easy switching and low brand loyalty and fast markets will play out. I hope that local LLMs will become very viable very soon.
Yeah I don’t think the models are meaningfully differentiated outside of very specific edge cases. I suspect this was the thinking behind OpenAI and Facebook and all trying to lean hard into presenting their chatbots as friends and romantic partners. If they can’t maintain a technical moat they can try to cultivate an emotional one.
Mathematics is hardly an edge case, but SOTA models differ wildly in their ability to write proofs for unsolved problems.
Models also differ wildly in tasks like decompilation for reverse engineering.
Also, so far, the only model I've found which can competently write PTX for SM100 CUDA devices is GPT-5.4pro, but I'm willing to admit that this is more of an edge case than the aforementioned.
AFAICT, the extent to which someone finds models interchangeable is inversely proportional to the novelty of their work.
Saw a comment here yesterday referencing the Attention Is All You Need paper title in a tongue in cheek way. Kinda fun to imagine the friend/romance angle is just a bunch of socially awkward folk at OpenAI misinterpreting the original paper
The big question is whether Apple can keep shipping new models constantly.
AFAIK the current model is on par with with Qwen-3-4B, which is from a year ago [0]. There's a big leap going from last year Qwen-3-4B to Qwen-3.5-4B or to Gemma 4.
Apple model is nice since you don't need to download anything else, but I'd rather use the latest model than to use a model from a year ago.
I’m not sure why that’s a question, it’s just a downloaded file. You can even watch it download separately when you enable Apple Intelligence (it’s not tied to OS updates from what I can tell).
Of course I imagine Apple is not going to be the fastest mover in this regard. I’m not even sure they believe the product will be widely impactful anymore and may keep it relegated to a small list of popular use cases like photo touch ups and quick questions to Siri. For me the most useful parts of Apple’s AI don’t even require me to enable Apple Intelligence.
Started using this earlier this week. I built a backtesting benchmark tool to compare a mix of frontier and open-source models on a fairly heavy data analysis workflow I’d been running in the cloud.
The task is basically predicting pricing and costs.
Apple’s model came out on top—best accuracy in 6 out of 10 cases in the backtest. That surprised me.
It also looks like it might be fast enough to take over the whole job. If I ran this on Sonnet, we’re talking thousands per month. With DeepSeek, it’s more like hundreds.
So far, the other local models I’ve tried on my 64GB M4 Max Studio haven’t been viable - either far too slow or not accurate enough. That said, I haven’t tested a huge range yet.
Saw this in an another thread previously and immediately installed it.
I have a new prompt to test LLMs much like simonw's pelican test.
"What is 9:30am Taiwan time in US, Pacific?" For some reason, the answers are quite inconsistent but all wrong.
./apfel "what is 9:30am Taiwan time in US, Pacific?"
Taiwan is 12 hours ahead of the Pacific Time Zone. Therefore, 9:30 AM Taiwan time would be 9:30 PM Pacific Time.
Taiwan is 13 hours ahead of the Pacific Time Zone. Therefore, 9:30 AM in Taiwan is 10:30 PM in the Pacific Time Zone.
Taiwan is in the China Standard Time (CST) zone, which is 12 hours ahead of the Pacific Standard Time (PST) zone. Therefore, 9:30 AM in Taiwan is 9:30 PM in the Pacific.
Taiwan is typically 11 hours ahead of the Pacific Time Zone. Therefore, 9:30 AM in Taiwan is 8:30 PM in the Pacific Time Zone.
Taiwan is 13 hours ahead of the Pacific Time Zone. Therefore, 9:30 AM in Taiwan is 10:30 PM the previous day in the Pacific Time Zone.
I'm a Linux user who wanted exactly this but for Linux — so I ended up building it myself. It's called TalkType, it runs Whisper locally for offline speech-to-text. The privacy angle was a big reason I went local from the start — I didn't want my voice being sent to anyone's server. Nice to see the same idea getting traction on Mac.
Tempted to write a grammarly-like underline engine that flags writing mistakes across all apps and browser. Fully private grammarly alternative without even bundling an LLM!
This doesn't feel truthful, it sounds like this tool is a hack that unlocks something. If I understand it correctly, it's using the same FoundationModels framework that powers Apple Intelligence, but for CLI and OpenAI compatible REST endpoint. Which is fine, just the marketing goes hard a bit.
> Runs on Neural Engine
Also unsure if this runs on ANE, when I tried Apple Intelligence I saw that it ran on the GPU (Metal).
Submitted a PR to prevent its installation on macos versions older than Tahoe(26), since I was able to install it on my older macos 15, but it aborted on execution.
The ATP episode where he talked about this was incredibly fascinating. Marco is such a role model to me - he has a complete immunity to fads and trends, and just does things the way he wants to. He adopts the 'new and cool' things only when they have a real benefit.
Just use the language you want when prompting it, like other LLMs?
% apfel "Gib mir ein Rezept für Currywurst."
Natürlich! Hier ist ein einfaches Rezept für Currywurst:
### Zutaten:
- **Für die Würste:**
- 500 g Bratwürste (z. B. Frankfurter Würste)
(note: clipped most of the reply, since I assume most of us here don't actually need an LLM-generated recipe)
Ok, then it might just be pretty bad at German Sie/Du. It manages to translate texts, but is unable to change between Du and Sie. interesting limitation I haven't encountered anywhere before. Even DeepL was able to swap between them.
Here is the chat I used to test this again to highlight this limitation:
you› write a short email to john asking for a good time to meet
ai› Subject: Let's Catch Up Soon!
Hi John,
I hope you're doing well! I'd love to grab a coffee or meet up for a chat sometime soon. When do you have some free time this week?
Looking forward to hearing from you!
Best,
[Your Name]
you› Please translate this to German
ai› Betreff: Lass uns bald treffen!
Hallo John,
Ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von Ihnen zu hören!
Liebe Grüße
[Dein Name]
you› Bitte mit Sie Form anstatt Du
ai› Betreff: Lass uns bald treffen!
Hallo John,
ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von Ihnen zu hören!
Liebe Grüße
[Dein Name]
you› Bitte alles mit Du anstatt Sie
ai› Betreff: Lass uns bald treffen!
Hallo John,
ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von dir zu hören!
Liebe Grüße
[Dein Name]
My config:
% apfel v0.6.13 — model info
├ model: apple-foundationmodel
├ on-device: true (always)
├ available: yes
├ context: 4096 tokens
├ languages: en, fr, zh, zh, ko, en, fr, de, tr, pt, it, pt, es, nl, da, nb, zh, es, es, ja, sv, vi, en
└ framework: FoundationModels (macOS 26+)
And for those who did know that and want to know more, the shift from apple - apfel and water -> wasser happened during the High German consonant shift.
Yeah I think you hit on the head a good way to use it though. I'm not on MacOS but KDE has a little tool called krunner[1] that lets you perform simple tasks from a small pop-up on your desktop. It would be cool if I could do slightly agentic things from there with a local model like ask what the capital of Austria is, or what's the current exchange rate between two currencies.
It's really handy for quick things like "what's the capital of country x" but for coding, I feel that it is severely limited. With such a small context it's (currently) not great for complicated things.
Local AIs are the future in times of limited resources. This could be the beginning of something big. I like that Apple opens up like this. Hopefully more to come.
To determine how many shopping days until Christmas, you'll need to know the current date and the date of Christmas. Christmas is on December 25th, so you can subtract the current date from this date to find the number of shopping days. If you provide the current date, I can help you calculate that.
apfel "How many shopping days until Christmas? Today is April 3rd."
To calculate the number of shopping days until Christmas, we first need to determine the date of Christmas this year.
Christmas in 2023 is on December 25th.
Starting from April 3rd, we calculate the days:
1. April: 30 days (from April 3rd to April 30th)
2. May: 31 days
3. June: 30 days
4. July: 31 days
5. August: 31 days
6. September: 30 days
7. October: 31 days
8. November: 30 days
Adding these up gives us:
30 + 31 + 30 + 31 + 31 + 30 + 31 + 30 = 252 days
Christmas is on December 25th, so we subtract 252 days from today, April 3rd:
252 - 252 = 0 days
So, there are 0 shopping days left until Christmas, but since Christmas is already past, there are no shopping days left.
This isn't a good test for any model since LLMs can't math (even though frontier models can sometimes correctly simulate mathing), which is why one would always use a tool for this.
Just discovered iOS shortcuts has a native action called “use model” that lets you use local, Apple cloud, or ChatGPT— before that I would have agreed with the author about being locked behind Siri (natively)
Would really love to see a web api standard for on device llms. This could get us closer. Some in-browser language model usage could be very powerful. In the interim maybe a little protocol spec + a discovery protocol used with browser plugins, web apps could detect and interface with on-device llms making it universally available.
As an experiment I built a prototype chatbot app that uses the built-in LLM. It’s got a small context window, but is surprisingly capable and has tool-calling support. Without too much effort I was able to get it to fetch weather data, fetch and summarise emails, read and write reminders and calendar events.
Just a small thing about the website: your examples shift all the elements below it on mobile when changing, making it jump randomly when trying to read.
I asked it to write a python program to find the intersection of a line and a circle and it one shot it. Asking for a Swift version worked as well.
Rust version goofed with lines of comments and code intertwined.
I asked what highway connects two given cities that it got wrong but would have been correct if I asked for two other cities.
I asked it to save answers to a file or notes but no burrito
Asked for an ascii drawing of a race car. Not sure what that thing was it returned.
Wow - the model really hallucinates without hesitation. I asked a number of "What do you know about [person, company, etc]?" questions and rather than realizing that it didn't know about them, it just made up an answer for every one of them.
Does the local LLM have access to personal information from the Apple account associated with the logged-in user? Maybe through a RAG pipeline or similar? Just curious if there are any risks associated with exposing this in a way that could be exploited via CORS or through another rogue app querying it locally.
no. the on device foundationmodels framework that apfel uses does not have access to personal information from the apple account. the model is a bare language model with no built in personal data access.
apple does have an on device rag pipeline called the semantic index that feeds personal data like contacts emails calendar and photos into the model context but this is only available to apples own first party features like siri and system summaries.
it is not exposed through the foundationmodels api.
1. Hugely non-deterministic: repeat queries give vastly different responses. 2. Often returns incorrect and inconsistent results even for mathematical queries. 3. Often the responses include unwanted highlighting or presentation markup. 4. Defaults to German decimal notation.
On a similar bent, I recently discovered Handy (cross-platform) which is very well implemented local voice input: https://handy.computer/ ... serious finger saver and ideal for LLM conversations
Thanks, tried it, but it's crashes on clicking the microphone icon. Default `make install` for some reason tries to install it to /usr, I changed that and after torturing more mature coding LLMs for 20 minutes, made it running with mic/sound.
The mic button requires clicking to transcribe and start listening again, and default voice is low-quality (I assume it can be configured).
In general I'm looking for a way to try the on-device hands-free voice mode.
If you’re looking into small models for tiny local tasks, you should try Qwen coder 0,5B. It’s more of an experiment, but it can output decent functions given the right context instructions.
So… a prompt? I’m not on my laptop but I hooked it to cmp.nvim, gave it a short situational prompt, +- 10 lines, and started typing. Not anywhere near usable but with a little effort you can get something ok for repetitive tasks. Maybe something like spotting one specific code smell pattern. The advantage is the ridiculous T/s you get
It’s a nice LLM because it seems fairly decent and it loads instantly and uses the CPU neural engine. The GPU is faster but when I run bigger LLMs on the GPU the normally very cool M series Mac becomes a lap roaster.
It’s a small LLM though. Seems decent but it’s also been safety trained to a somewhat comical degree. It will balk over safety at requests that are in fact quite banal.
This is great. I think Apple bought Kuzu, a in memory graph database in late 2025 to support RAG in combine with their FM like this. Even with such as small model, a comprehensive context of our personal data in graph RAG would be sufficient for a PA system. Do we know if we can have access to this RAG data?
Interesting approach. I've been building MCP tools for AI agents recently and the local-first model makes a lot of sense for privacy-sensitive use cases.
Does Apfel support MCP tool calling?
the 2 hard limits of Appel Intelligence Foundation Model and therefor apfel is the 4k token context window and the super hard guardrails (the model prefers to tell you nothing before it tells you something wrong ie ask it to describe a color)
parsing logfiles line by line, sure
parsing a whole logfile, well it must be tiny, logfile hardly ever are
Which Apple silicon support this and how good are the models available to different generations? I’m assume an M4 can and will run bigger models than an M2
This is pretty cool. My bet is that we have more LLMs running locally when its possible, either thru "better hardware as default" or some new tech that can run the models on commodity hardware (like apple silicon / equivalent PC setup).
I have been using Apple’s built-in system LLM model for the last 7 or 8 months. I like the feature that if it needs to, it occasionally uses a more powerful secure private cloud model. I also write my own app to wrap it.
wow, had a very similar idea, but smaller in scope. (focusing on shell script usage) I've documented my own journey over here https://github.com/greg76/aii
dyld[71398]: Library not loaded: /System/Library/Frameworks/FoundationModels.framework/Versions/A/FoundationModels
Referenced from: <32818E2F-CB45-3506-A35B-AAF8BDDFFFCE> /opt/homebrew/Cellar/apfel/0.6.25/bin/apfel (built for macOS 26.0 which is newer than running OS)
Reason: tried: '/System/Library/Frameworks/FoundationModels.framework/Versions/A/FoundationModels' (no such file), '/System/Volumes/Preboot/Cryptexes/OS/System/Library/Frameworks/FoundationModels.framework/Versions/A/FoundationModels' (no such file), '/System/Library/Frameworks/FoundationModels.framework/Versions/A/FoundationModels' (no such file, not in dyld cache)
MacBook Neo forced me to finally make the jump, and it turns out that I, much like the engineers at Apple, don't really care about the spit and finish anymore. Third-party applications handle everything else. Also, I was happy to find that Divvy still runs just fine under Rosetta.
Looks like a nice wrapper around the APIs. Extremely oversold landing page, very marketing heavy for what it is. You can actually make nice looking landing pages that are about 10% the size of this and more straightforward, rather than some mimicry of a SaaS that's trying desperately to sell you something. Makes it easier for you to review the content for factuality too, and heck you couldn't even take ownership of some of the voice.
Hard to know what to do with this. I'm interested in the project and know others who would be, but I feel like shit after being slopped on by a landing page and I don't wish to slop on my friends by sharing it with them. I suppose the github link is indeed significantly better, I'll share that.
I certainly don't feel entitled to anybody's effort, and I appreciate the project itself!
It's just these flashy LLM-generated webpages are really misleading.
It's filled with attention-grabbing LLM-filler. It presents itself as high-effort content, but in reality is unclear if it's accurate, unclear how much the author even put into READING the content they're asking me to read, and overall just really unclear what parts of the page are important to the author and what are just stuff the LLM inserted into a marketing card to make it look like a SaaS sales pitch.
If you dislike the slop on the website so much you might also not like GitHub that much. It shows this software has been heavily co-written by Claude code
This is great. A few questions come to mind, I need to go look up. Is the model an OpenAI one or home grown for Apple. And can I still use it if Siri is disabled?
I find the branding to be a little odd. Like, it should be a github page with a README that says "here's how to use this." Like, the full explanation of this project is right there in the HN title: "The free AI already on your Mac."
I guess LLMs have made it too simple to instantly build startup landing page slop, which causes this? Like, do we need to see the github star count chart? Do we need all the buzzwords and stuff? You'd think this was a startup trying to get a billion dollar evaluation. It feels disingenuous.
> Referenced from: <32818E2F-CB45-3506-A35B-AAF8BDDFFFCE> /opt/homebrew/Cellar/apfel/0.6.25/bin/apfel (built for macOS 26.0 which is newer than running OS)
This actually looks really neat. I'll have to bookmark this for whenever I'm dragged kicking and screaming into the abomination that is "Tahoe."
I like the approach of running everything locally. I'm strongly of the opinion that the privacy angle for local models is going to keep getting stronger and more relevant. The amount of articles that come out about accidents happening because of people handing too much context to cloud models the more self reinforcing this will become.
It's only half of the solution though. If the models are trained in a closed way, they can prioritize values encoded during training even if that's not what you want (example: ask the open Chinese models about Tiananmen). It's not beyond imagining that these models would e.g. try to send your data to authorities or advertisers when their training says so, even if you run them locally.
So the full solution would be models trained in an open verifiable way and running locally.
All it would take is for one person to catch the model doing this and the reputation of the model and the company would be destroyed irrevocably.
The model is only generating tokens without touching the network at all, right? How would it send data away?
Theoretically, by taking the opportunity to inject an exfiltration mechanism if you ask it to write code for you
Lots of people I know run models in "yolo" mode or the equivalent as well, which means it could just invoke curl or telnet to exfiltrate data.
Another angle is when you're passing untrusted content to the AI service, e.g. anything from using it to crawl websites to spam-detection on new forum user posts.
You can trigger the the service's ToS violation or worse, get tipped off to law enforcement for something you didn't even write.
local is best for privacy, but i personally think you don't need to go local.
anthropic, google, openai etc, decided that their consumer ai plans would not be private. partly to collect training data, the other half to employ moderators to review user activity for safety.
we trust that human moderators will not review and flag our icloud docs, onedrive or gmail, or aggregate such documents into training data for llms. it became the norm that an llm is somehow not private. it became a norm that you can't opt out of training, even on paid plans (see meta and google); or if you can opt out of training, you can't opt out of moderation.
cloud models with a zero retention privacy policy are private enough for almost everyone, the subscriptions, google search, ai search engines are either 'buying' your digital life or covering themselves for legal reasons.
you can and should have private cloud services, and if legal agreement is not enough, cryptographic attestation is already used in compute, with AWS nitro enclaves and other providers.
> i personally think you don't need to go local.
I personally think everyone should default to using local resources. Cloud resources should only be used for expansion and be relatively bursty rather than the default.
For about two years I experimented with writing local apps using local LLMs, but I often had to blend in a commercial web search API to make my little experiments useful.
I pay $13/month for Proton’s Lumo+ private chat LLM that contains an excellent built-in web search tool. I use it for everything non-technical, even just simple searching for local businesses, etc.
As an enthusiastic reader of books like Privacy is Power and Surveillance Capitalism, it feels good to have a private tool that is ready at hand.
do you have any provider recommendations? I've experimented with this on runpod serverless, but I've been meaning to dig deeper before I feel comfortable with personal data.
I saw a service named Phala, which claims to be actually no-knowledge to server side (I think). It was significantly more expensive, but interesting to see it's out there. My thought was escaping the data-collection-hungry consumer models was a big win.
> anthropic, google, openai etc, decided that their consumer ai plans would not be private. partly to collect training data, the other half to employ moderators to review user activity for safety.
That's two halves of "why", sure.
Another interesting half would be that those companies have US military officers on their boards, and LLMs are the ultimate voluntary data collection platform, even better trojan horses than smartphones.
Yet another "half" could be how much enterprise value might be found by datamining for a minute or two... may I suggest reading a couple of Martha Wells books.
That's the way things have to go. Business risk is too high having everything ran over exposed networks.
what i say about this, is that an llm is just a big file, there is nothing 'not private' about it.
if you are happy with off-prem then the llm is ok too, if you need on-prem this is when you will need local.
> an llm is just a big file, there is nothing 'not private' about it.
The private thing is the prompt.
But also, a local LLM opens up the possibility of agentic workflows that don't have to touch the Internet.
The other thing, is encrypted inferencing a thing/service currently? I want to run my own models locally just because if I'm going to be chatting to it about my day to day life why send it to a server in plaintext.
encrypted inferencing, meaning homomorphic encryption: no, it's not solved.
cryptographic confirmation of zero knowledge: yes.
the latter, based on trust in the hardware manufacturer and their root ca. so, encrypted if you trust intel/nvidia to sign it.
there are a few services, phala, tinfoil, near ai, redpill is an aggregator of those
> I like the approach of running everything locally. I'm strongly of the opinion that the privacy angle for local models is going to keep getting stronger and more relevant.
In HN circles perhaps. Average Joes don’t care.
I bet if you clearly explained the benefits and tradeoffs, and then gave them the choice, Average Joes would care.
They generally do care, but not enough to change what they do or to do without something they use, like social media.
So many people I know say “I only use Signal to talk to you”, it’s like I’m the awkward one for not using Facebook.
I’ve seen several projects like this that offer a network server with access to these Apple models. The danger is when they expose that, even on a loop port, to every other application on your system, including the browser. Random webpages are now shipping with JavaScript that will post to that port. Same-origin restrictions will stop data flow back to the webpage, but that doesn’t stop them from issuing commands to make changes.
Some such projects use CORS to allow read back as well. I haven’t read Apfel’s code yet, but I’m registering the experiment before performing it.
They offer it as an option but default it to false! This is still a --footgun option but it’s the least unsafe version I’ve seen yet! Well done, Apfel authors.
thx for the report - a totally valid attack vector i was not aware of before, should be fixed https://github.com/Arthur-Ficial/apfel/releases/tag/v0.6.23 - see also new https://github.com/Arthur-Ficial/apfel/blob/main/docs/server...
I don’t think many browsers will allow posting to 127.0.0.1 from a random website. What’s the threat model here?
Restricting such access it is still a work in progress: https://wicg.github.io/local-network-access/
I think any browser will allow it but not allow data read back.
FWIW this was the status quo (webpage could ping arbitrary ports but not read data, even with CORS protections) - but it is changing.
This is partially in response to https://localmess.github.io/ where Meta and Yandex pixel JS in websites would ping a localhost server run by their Android apps as a workaround to third-party cookie limits.
Chrome 142 launched a permission dialog: https://developer.chrome.com/blog/local-network-access
Edge 140 followed suit: https://support.microsoft.com/en-us/topic/control-a-website-...
And Firefox is in progress as well, though I couldn't find a clear announcement about rollout status: https://fosdem.org/2026/schedule/event/QCSKWL-firefox-local-...
So things are getting better! But there was a scarily long time where a rogue JS script could try to blindly poke at localhost servers with crafty payloads, hoping to find a common vulnerability and gain RCE or trigger exfiltration of data via other channels. I wouldn't be surprised if this had been used in the wild.
There is a CORS preflight check for POST requests that don't use form-encoding. It would be somewhat surprising if these weren't using JSON (though it wouldn't be that surprising if they were parsing submitted JSON instead of actually checking the MIME-type which would probably be bad anwyay)
Isn't there a CORS preflight check for this? In most cases. I guess you could fashion an OG form to post form fields. But openai is probably a JSON body only.
The default scenario should be secure. If the local site sends permissive CORS headers bets may be off. I would need to check but https->http may be a blocker too even in that case. Unless the attack site is http.
Keep seeing similar mistakes with vibe coded AI & MCP projects. Even experienced engineers seem oblivious to this attack vector
Noting that there's an option to require a Bearer token to the API
Now this is a development I like.
With the Claude bug, or so it is known, burning through tokens at record speed, I gave alternative models a try and they're mostly ... interchangeable. I don't know how easy switching and low brand loyalty and fast markets will play out. I hope that local LLMs will become very viable very soon.
Yeah I don’t think the models are meaningfully differentiated outside of very specific edge cases. I suspect this was the thinking behind OpenAI and Facebook and all trying to lean hard into presenting their chatbots as friends and romantic partners. If they can’t maintain a technical moat they can try to cultivate an emotional one.
> very specific edge cases
Mathematics is hardly an edge case, but SOTA models differ wildly in their ability to write proofs for unsolved problems.
Models also differ wildly in tasks like decompilation for reverse engineering.
Also, so far, the only model I've found which can competently write PTX for SM100 CUDA devices is GPT-5.4pro, but I'm willing to admit that this is more of an edge case than the aforementioned.
AFAICT, the extent to which someone finds models interchangeable is inversely proportional to the novelty of their work.
Saw a comment here yesterday referencing the Attention Is All You Need paper title in a tongue in cheek way. Kinda fun to imagine the friend/romance angle is just a bunch of socially awkward folk at OpenAI misinterpreting the original paper
The big question is whether Apple can keep shipping new models constantly.
AFAIK the current model is on par with with Qwen-3-4B, which is from a year ago [0]. There's a big leap going from last year Qwen-3-4B to Qwen-3.5-4B or to Gemma 4.
Apple model is nice since you don't need to download anything else, but I'd rather use the latest model than to use a model from a year ago.
https://machinelearning.apple.com/research/apple-foundation-...
I’m not sure why that’s a question, it’s just a downloaded file. You can even watch it download separately when you enable Apple Intelligence (it’s not tied to OS updates from what I can tell).
Of course I imagine Apple is not going to be the fastest mover in this regard. I’m not even sure they believe the product will be widely impactful anymore and may keep it relegated to a small list of popular use cases like photo touch ups and quick questions to Siri. For me the most useful parts of Apple’s AI don’t even require me to enable Apple Intelligence.
Started using this earlier this week. I built a backtesting benchmark tool to compare a mix of frontier and open-source models on a fairly heavy data analysis workflow I’d been running in the cloud.
The task is basically predicting pricing and costs.
Apple’s model came out on top—best accuracy in 6 out of 10 cases in the backtest. That surprised me.
It also looks like it might be fast enough to take over the whole job. If I ran this on Sonnet, we’re talking thousands per month. With DeepSeek, it’s more like hundreds.
So far, the other local models I’ve tried on my 64GB M4 Max Studio haven’t been viable - either far too slow or not accurate enough. That said, I haven’t tested a huge range yet.
Digging into this, found Apple’s release notes for the Foundation Model Service
https://developer.apple.com/documentation/Updates/Foundation...
They released an official python SDK in March 2026:
https://github.com/apple/python-apple-fm-sdk
Saw this in an another thread previously and immediately installed it.
I have a new prompt to test LLMs much like simonw's pelican test.
"What is 9:30am Taiwan time in US, Pacific?" For some reason, the answers are quite inconsistent but all wrong.
I'm a Linux user who wanted exactly this but for Linux — so I ended up building it myself. It's called TalkType, it runs Whisper locally for offline speech-to-text. The privacy angle was a big reason I went local from the start — I didn't want my voice being sent to anyone's server. Nice to see the same idea getting traction on Mac.
Tempted to write a grammarly-like underline engine that flags writing mistakes across all apps and browser. Fully private grammarly alternative without even bundling an LLM!
That's a great idea. I would be very interested in using it of someone builds it.
> Apple locked it behind Siri. apfel sets it free
This doesn't feel truthful, it sounds like this tool is a hack that unlocks something. If I understand it correctly, it's using the same FoundationModels framework that powers Apple Intelligence, but for CLI and OpenAI compatible REST endpoint. Which is fine, just the marketing goes hard a bit.
> Runs on Neural Engine
Also unsure if this runs on ANE, when I tried Apple Intelligence I saw that it ran on the GPU (Metal).
AFM models are very impressive, but they’re not made for conversation, so keep your expectations down in chat mode.
Thank you for making it open source!
Submitted a PR to prevent its installation on macos versions older than Tahoe(26), since I was able to install it on my older macos 15, but it aborted on execution.
https://github.com/Arthur-Ficial/homebrew-tap/pull/1
It’s a very small model but I’ve been playing with it for some time now I’m impressed. Have we been sleeping on Apple’s models?
Imagine they baked Qwen 3.5 level stuff into the OS. Wow that’d be cool.
Apparently the Overcast guy build a beowulf cluster of Mac minis to use the Apple transcription service.
https://www.linkedin.com/posts/nathangathright_marco-arment-...
The ATP episode where he talked about this was incredibly fascinating. Marco is such a role model to me - he has a complete immunity to fads and trends, and just does things the way he wants to. He adopts the 'new and cool' things only when they have a real benefit.
I largely subscribe to the use boring tech ethos ... but php? come on man.
and yet... successful people have used it to build really successful things: Facebook, Tumblr (I think), the things Marco's been involved with.
I just dunno outside of meta should we really be pushing php with all its flaws? or is it still flawed and I need to update my priors?
There’s something major to be said for going to war with the tools you have.
And living with decisions made 15 years ago may be much more successful than trying to change horses mid-stream.
Yeah heard him talk about that. 48 or so 16GB m4 Mac minis. Insane. The Beowulf lives
The vision models and OCR are SUPER
For small tasks this seems perfect. However it being limited to English from what I can tell is quite a downsite for me.
It can work in other languages?
Just use the language you want when prompting it, like other LLMs? (note: clipped most of the reply, since I assume most of us here don't actually need an LLM-generated recipe)Ok, then it might just be pretty bad at German Sie/Du. It manages to translate texts, but is unable to change between Du and Sie. interesting limitation I haven't encountered anywhere before. Even DeepL was able to swap between them.
Here is the chat I used to test this again to highlight this limitation:
you› write a short email to john asking for a good time to meet ai› Subject: Let's Catch Up Soon!
Hi John,
I hope you're doing well! I'd love to grab a coffee or meet up for a chat sometime soon. When do you have some free time this week?
Looking forward to hearing from you!
Best, [Your Name]
you› Please translate this to German ai› Betreff: Lass uns bald treffen!
Hallo John,
Ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von Ihnen zu hören!
Liebe Grüße [Dein Name]
you› Bitte mit Sie Form anstatt Du ai› Betreff: Lass uns bald treffen!
Hallo John,
ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von Ihnen zu hören!
Liebe Grüße [Dein Name]
you› Bitte alles mit Du anstatt Sie ai› Betreff: Lass uns bald treffen!
Hallo John,
ich hoffe, es geht dir gut! Ich würde gerne Kaffee trinken oder uns für ein Gespräch treffen. Wann haben Sie in dieser Woche freie Zeit?
Ich freue mich darauf, von dir zu hören!
Liebe Grüße [Dein Name]
My config: % apfel v0.6.13 — model info ├ model: apple-foundationmodel ├ on-device: true (always) ├ available: yes ├ context: 4096 tokens ├ languages: en, fr, zh, zh, ko, en, fr, de, tr, pt, it, pt, es, nl, da, nb, zh, es, es, ja, sv, vi, en └ framework: FoundationModels (macOS 26+)
Nice! The example should imo say
apfel -o json "Translate to German: apple" | jq .content
Thx. Good point. Done.
For those who don't know, 'Apfel' is the German word for Apple.
And for those who did know that and want to know more, the shift from apple - apfel and water -> wasser happened during the High German consonant shift.
https://en.wikipedia.org/wiki/High_German_consonant_shift
I hacked this together last fall to let you use Apple Foundation Models with llm: https://github.com/btucker/llm-apple . To enable that I built python bindings with Claude Code: https://github.com/btucker/apple-foundation-models-py
Unfortunately, I found the small context window makes the utility pretty limited.
Yeah I think you hit on the head a good way to use it though. I'm not on MacOS but KDE has a little tool called krunner[1] that lets you perform simple tasks from a small pop-up on your desktop. It would be cool if I could do slightly agentic things from there with a local model like ask what the capital of Austria is, or what's the current exchange rate between two currencies.
Then save the heavy lifting for the big boys.
[1] https://userbase.kde.org/Plasma/Krunner
I've used Alfred[1] for many years on macOS, and more recently the PowerToys Command Palette[2] on Windows.
Both are easily extensible, so you could certainly wire either up to the local (or hosted) LLM of your choice.
Side note: I just noticed the extension example in the Command Palette demo reel searches HN.
[1] https://www.alfredapp.com/
[2] https://learn.microsoft.com/en-us/windows/powertoys/command-...
This is similar to something I was playing around with last month-- basically just a CLI for accessing the foundational models.
https://github.com/ehamiter/afm
It's really handy for quick things like "what's the capital of country x" but for coding, I feel that it is severely limited. With such a small context it's (currently) not great for complicated things.
Local AIs are the future in times of limited resources. This could be the beginning of something big. I like that Apple opens up like this. Hopefully more to come.
completely agree.
“Already on your Mac” is a strong hook.
Curious where you see local-first AI breaking down compared to API-based workflows.
Christmas in 2023 is on December 25th.
Starting from April 3rd, we calculate the days:
1. April: 30 days (from April 3rd to April 30th) 2. May: 31 days 3. June: 30 days 4. July: 31 days 5. August: 31 days 6. September: 30 days 7. October: 31 days 8. November: 30 days
Adding these up gives us:
30 + 31 + 30 + 31 + 31 + 30 + 31 + 30 = 252 days
Christmas is on December 25th, so we subtract 252 days from today, April 3rd:
252 - 252 = 0 days
So, there are 0 shopping days left until Christmas, but since Christmas is already past, there are no shopping days left.
Yep, feels like Siri.
This isn't a good test for any model since LLMs can't math (even though frontier models can sometimes correctly simulate mathing), which is why one would always use a tool for this.
Real experience I've had:
"Text Carol bring me a glass of water please"
"I'm sorry, I don't see a 'Carol Bring' in your contacts"
Almost about to try it until I saw this. If it's Siri the Silly don't even make up for the opportunity cost.
Just discovered iOS shortcuts has a native action called “use model” that lets you use local, Apple cloud, or ChatGPT— before that I would have agreed with the author about being locked behind Siri (natively)
Would really love to see a web api standard for on device llms. This could get us closer. Some in-browser language model usage could be very powerful. In the interim maybe a little protocol spec + a discovery protocol used with browser plugins, web apps could detect and interface with on-device llms making it universally available.
https://webmachinelearning.github.io/prompt-api/
Already in Chrome as an origin trial: https://developer.chrome.com/docs/ai/prompt-api
As an experiment I built a prototype chatbot app that uses the built-in LLM. It’s got a small context window, but is surprisingly capable and has tool-calling support. Without too much effort I was able to get it to fetch weather data, fetch and summarise emails, read and write reminders and calendar events.
Just a small thing about the website: your examples shift all the elements below it on mobile when changing, making it jump randomly when trying to read.
I asked it to write a python program to find the intersection of a line and a circle and it one shot it. Asking for a Swift version worked as well. Rust version goofed with lines of comments and code intertwined.
I asked what highway connects two given cities that it got wrong but would have been correct if I asked for two other cities.
I asked it to save answers to a file or notes but no burrito
Asked for an ascii drawing of a race car. Not sure what that thing was it returned.
Wow - the model really hallucinates without hesitation. I asked a number of "What do you know about [person, company, etc]?" questions and rather than realizing that it didn't know about them, it just made up an answer for every one of them.
Does the local LLM have access to personal information from the Apple account associated with the logged-in user? Maybe through a RAG pipeline or similar? Just curious if there are any risks associated with exposing this in a way that could be exploited via CORS or through another rogue app querying it locally.
no. the on device foundationmodels framework that apfel uses does not have access to personal information from the apple account. the model is a bare language model with no built in personal data access.
apple does have an on device rag pipeline called the semantic index that feeds personal data like contacts emails calendar and photos into the model context but this is only available to apples own first party features like siri and system summaries.
it is not exposed through the foundationmodels api.
1. Hugely non-deterministic: repeat queries give vastly different responses. 2. Often returns incorrect and inconsistent results even for mathematical queries. 3. Often the responses include unwanted highlighting or presentation markup. 4. Defaults to German decimal notation.
On a similar bent, I recently discovered Handy (cross-platform) which is very well implemented local voice input: https://handy.computer/ ... serious finger saver and ideal for LLM conversations
What's the easiest way to use it with on-device voice model for voice chat?
https://github.com/Arthur-Ficial/apfel-gui uses on-device speech-to-text and text-to-speech
Thanks, tried it, but it's crashes on clicking the microphone icon. Default `make install` for some reason tries to install it to /usr, I changed that and after torturing more mature coding LLMs for 20 minutes, made it running with mic/sound.
The mic button requires clicking to transcribe and start listening again, and default voice is low-quality (I assume it can be configured).
In general I'm looking for a way to try the on-device hands-free voice mode.
https://handy.computer
Anyone tried using this as a sub-agent for a more capable model like Claude/Codex?
If you’re looking into small models for tiny local tasks, you should try Qwen coder 0,5B. It’s more of an experiment, but it can output decent functions given the right context instructions.
> [Qwen coder 0,5B] can output decent functions given the right context instructions
Can you share a working example?
So… a prompt? I’m not on my laptop but I hooked it to cmp.nvim, gave it a short situational prompt, +- 10 lines, and started typing. Not anywhere near usable but with a little effort you can get something ok for repetitive tasks. Maybe something like spotting one specific code smell pattern. The advantage is the ridiculous T/s you get
The combined (input/output) context window length is 4K. Claude would blow through that even when trying to read and summarize a small file.
With a small/minimalistic harness like Pi maybe it works well?
I was thinking about the other way: Could you use this in front of Claude to summarize inputs and so reduce your token counts?
project started with
trying to run openclaw with it in ultra token saving mode, did totally not work.
great for shell scripts though (my major use case now)
BoltAI also does this, but a CLI tool is nice.
It’s a nice LLM because it seems fairly decent and it loads instantly and uses the CPU neural engine. The GPU is faster but when I run bigger LLMs on the GPU the normally very cool M series Mac becomes a lap roaster.
It’s a small LLM though. Seems decent but it’s also been safety trained to a somewhat comical degree. It will balk over safety at requests that are in fact quite banal.
This is great. I think Apple bought Kuzu, a in memory graph database in late 2025 to support RAG in combine with their FM like this. Even with such as small model, a comprehensive context of our personal data in graph RAG would be sufficient for a PA system. Do we know if we can have access to this RAG data?
Interesting approach. I've been building MCP tools for AI agents recently and the local-first model makes a lot of sense for privacy-sensitive use cases. Does Apfel support MCP tool calling?
Really like demo cli tools description. Are they limited by the context window as well? What’s your experience with log file sizes?
the 2 hard limits of Appel Intelligence Foundation Model and therefor apfel is the 4k token context window and the super hard guardrails (the model prefers to tell you nothing before it tells you something wrong ie ask it to describe a color)
parsing logfiles line by line, sure
parsing a whole logfile, well it must be tiny, logfile hardly ever are
the model prefers to tell you nothing before it tells you something wrong
If all LLMs did this, people would trust them more.
Which Apple silicon support this and how good are the models available to different generations? I’m assume an M4 can and will run bigger models than an M2
I like the idea and the clarity to explain the usage, my question would be: what kind of tasks it would be useful for?
Making a sentence out of a json
This is pretty cool. My bet is that we have more LLMs running locally when its possible, either thru "better hardware as default" or some new tech that can run the models on commodity hardware (like apple silicon / equivalent PC setup).
How does this model compare against other local models like Qwen run through LMStudio?
I have been using Apple’s built-in system LLM model for the last 7 or 8 months. I like the feature that if it needs to, it occasionally uses a more powerful secure private cloud model. I also write my own app to wrap it.
But by access to the private compute right?
wow, had a very similar idea, but smaller in scope. (focusing on shell script usage) I've documented my own journey over here https://github.com/greg76/aii
Any know if these only installed on Tahoe? I'm running Sequoia still and get an error about model not found.
> Apple Silicon Mac, macOS 26 Tahoe or newer, Apple Intelligence enabled
Yes, the model ships with Tahoe, not previous versions.
I too would love to try this for simple prompts but won’t be updating past Sequoia for the foreseeable future.
Same. What a disaster Tahoe is.
How much storage does it take up?
4mb download, after install about 15mb, model is already on your mac with mac os x tahoe
It's not, unless you've installed it by turning the feature on in settings. That's where the heavy download will be.
I want access to the LLM xcode uses for swift auto completions
Awesome idea. You should launch this on Buildfeed.co.
Is this for Tahoe only? I’m still clutching onto Sequoia
Yeah seems to need Tahoe (I'm on Sequoia):
Yes, it says on that page that it uses Apple Intelligence from Tahoe. I'm also hanging onto Sequoia, though I'm ready to make the leap any time here.
MacBook Neo forced me to finally make the jump, and it turns out that I, much like the engineers at Apple, don't really care about the spit and finish anymore. Third-party applications handle everything else. Also, I was happy to find that Divvy still runs just fine under Rosetta.
love the simple website and typography. AI design or you? tasteful and fast animations. nice work and thanks for sharing!
Looks like a nice wrapper around the APIs. Extremely oversold landing page, very marketing heavy for what it is. You can actually make nice looking landing pages that are about 10% the size of this and more straightforward, rather than some mimicry of a SaaS that's trying desperately to sell you something. Makes it easier for you to review the content for factuality too, and heck you couldn't even take ownership of some of the voice.
Hard to know what to do with this. I'm interested in the project and know others who would be, but I feel like shit after being slopped on by a landing page and I don't wish to slop on my friends by sharing it with them. I suppose the github link is indeed significantly better, I'll share that.
It's absolutely free and open source, no need to bash it like this.
I certainly don't feel entitled to anybody's effort, and I appreciate the project itself!
It's just these flashy LLM-generated webpages are really misleading.
It's filled with attention-grabbing LLM-filler. It presents itself as high-effort content, but in reality is unclear if it's accurate, unclear how much the author even put into READING the content they're asking me to read, and overall just really unclear what parts of the page are important to the author and what are just stuff the LLM inserted into a marketing card to make it look like a SaaS sales pitch.
If you dislike the slop on the website so much you might also not like GitHub that much. It shows this software has been heavily co-written by Claude code
I'm aware of that! That's fine with me.
This is great. A few questions come to mind, I need to go look up. Is the model an OpenAI one or home grown for Apple. And can I still use it if Siri is disabled?
This is cool!
You have to enable Apple Intelligence so that's a hard no from me. I'll stick to LM Studio and gpt-oss/qwen. Very cool project though.
Interesting. How does this foundational model compares with other LLMs?
Read Austria as Australia and thought this as an April fool
Tahoe+ only
I find the branding to be a little odd. Like, it should be a github page with a README that says "here's how to use this." Like, the full explanation of this project is right there in the HN title: "The free AI already on your Mac."
I guess LLMs have made it too simple to instantly build startup landing page slop, which causes this? Like, do we need to see the github star count chart? Do we need all the buzzwords and stuff? You'd think this was a startup trying to get a billion dollar evaluation. It feels disingenuous.
Maybe I'm just being a hater.
Notes.app handles big notebooks without choking on storage?
Cool tool but I don't get why these websites make idiotic claims
> $0 cost
No kidding.
Why not just link the GH Github: https://github.com/Arthur-Ficial/apfel
He did?
https://news.ycombinator.com/item?id=47624647
A serious project would do the work to be delivered via the native homebrew repository, not a “selfhosted” one.
Isn't the whole idea of "home brew" to enable hackers and enthusiasts to easily share what they built?
Is this you signing up as a packager or
> Starting with macOS 26 (Tahoe), every Apple Silicon Mac includes a language model as part of Apple Intelligence.
So you have to put up with the low contrast buggy UI to use that.
> Referenced from: <32818E2F-CB45-3506-A35B-AAF8BDDFFFCE> /opt/homebrew/Cellar/apfel/0.6.25/bin/apfel (built for macOS 26.0 which is newer than running OS)
This actually looks really neat. I'll have to bookmark this for whenever I'm dragged kicking and screaming into the abomination that is "Tahoe."