> I think this is a game changer, because data privacy is a legitimate concern for many enterprise users.

Indeed. At work, we are experimenting with this. Using a cloud platform is a non-starter for data confidentiality reasons. On-premise is the way to go. Also, they’re not American, which helps.

> Btw, you can also run Mistral locally within the Docker model runner on a Mac.

True, but you can do that only with their open-weight models, right? They are very useful and work well, but their commercial models are bigger and hopefully better (I use some of their free models every day, but none of their commercial ones).

I also kind of don't understand how it seems everyone is using AI for coding. I haven't had a client yet which would have approved any external AI usage. So I basically use them as search engines on steroids, but code can't go directly in or out.

You might be able to get your clients to sign something to allow usage, but if you don't, as you say, it doesn't seem wise to vibe code for them. For two reasons:

1. A typical contract transfers the rights to the work. The ownership of AI generated code is legally a wee bit disputed. If you modify and refactor generated code heavily it's probably fine, but if you just accept AI generated code en masse, making your client think that you wrote it and it is therefore their copyright, that seems dangerous.

2. A typical contract or NDA also contains non disclosure, i.e. you can't share confidential information, e.g. code (including code you _just_ wrote, due to #1) with external parties or the general public willy nilly. Whether any terms of service assurances from OpenAI or Anthropic that your model inputs and outputs will probably not be used for training are legally sufficient, I have doubts.

IANAL, and _perhaps_ I'm wrong about one or both of these, in one or more countries, but by and large I'd say the risk is not worth the benefit.

I mostly use third party LLMs like I would StackOverflow: Don't post company code there verbatim, make an isolated example. And also don't paste from SO verbatim. I tried other ways of using LLMs for programming a few times in personal projects and can't say I worry about lower productivity with these limitations. YMMV.

(All this also generally goes for employees with typical employment contracts: It's probably a contract violation.)

Yes these are indeed the points. I don't really care too much, it would make me a bit more efficient but I'm billing by the hour anyway so I'm completely fine playing by the book.

Not sure I can agree with the "I'm billing by the hour" part.

I mean sure, but I think of my little agency providing value, for a price. Clients have budgets, they have limited benefits from any software they build, and in order to be competitive against other agencies or their internal teams, overall, I feel we need to provide a good bang for buck.

But since it's not all that much about typing in code, and since even that activity isn't all that sped up by LLMs, not if quality and stability matters, I would still agree that it's completely fine.

Yes, it's important of course that I'm efficient, and I am. But my coding speed isn't the main differentiating factor why clients like me.

I meant that I don't care enough to spearhead and drive this effort within the client orgs. They have their own processes, and internal employees would surely also like to use AI, so maybe they'll get there eventually. And meanwhile I'll just use it in the approved ways.

What about 10 years ago when we all copied code from SO? Did we worry about copyright then? Maybe we did and I don’t recall.

It's roughly the same, legally, and I was well aware of that.

Legally speaking, you also want to be careful about your dependencies and their licenses, a company that's afraid to get sued usually goes to quite some lengths to ensure they play this stuff safe. A lot of smaller companies and startups don't know or don't care.

From a professional ethics perspective, personally, I don't want to put my clients in that position unless they consciously decide they want that. They hire professionals not just to get work done they fully understand, but to a large part to have someone who tells them what they don't know.

You raise a good point. It was kinda gray in the SO days. You almost always had to change something to get your code to work. But at lot of LLM's can spit out code that you can just paste in. And, I guess maybe the tests all pass, but if it goes wrong, you, the coder probably don't know where it went wrong. But if you'd written it all yourself, you could probably guess.

I'm still sorting all this stuff out personally. I like LLM's when I work in an area I know well. But vibing in areas of technology that I don't know well just feels weird.

SO seems different because the author of the post is republishing it. If they are republishing copyrighted material without notice, it seems like the SO author is the one in violation of copyright.

In the LLM case, I think it’s more of an open question whether the LLM output is republishing the copyrighted content without notice, or simply providing access to copyrighted content. I think the former would put the LLM provider in hot water, while the latter would put the user in hot water.

“We” took care to not copy it verbatim (it’s the concrete code form that is copyrighted, not the algorithm), and depending on jurisdiction there is the concept of https://en.wikipedia.org/wiki/Threshold_of_originality in copyright law, which short code snippets on Stack Overflow typically don’t meet.

Nobody is seriously disputing the ownership of AI generated code. A serious dispute would be a considerable, concerted effort to stop AI code generation in any jurisdiction, that provides a contrast to the enormous, ongoing efforts by multiple large players with eye-watering investments to make code generation bigger and better.

Note, that this is not a statement about the fairness or morality of LLM building, but to think that the legality of AI code generation is something to reasonably worry about, is betting against multiple large players and their hundreds of billions of dollars in investment right now, and that likely puts you in a bad spot in reality.

> Nobody is seriously disputing the ownership of AI generated code

From what I've been following it seems very likely that, at least in the US, AI-generated anything can't actually be copyrighted and thus can't have ownership at all! The legal implications of this are yet to percolate through the system though.

Only if that interpretation lasts despite likely intense lobbying to the contrary.

this is "Kool-aid" from the supply side of LLMs for coding IMO. Plenty of people are plenty upset about the capture of code at Github corral, fed into BigCorp$ training systems.

parent statement reminds me of smug French in a castle north of London circa 1200, with furious locals standing outside the gates, dressed in rags with farm tools as weapons. One well-equipped tower guard says to another "no one is seriously disputing the administration of these lands"

I think the comparison falls flat, but it's actually really funny. I'll keep it in mind.

Your mother was a hamster and your father smelt of elderberries?

Other forms of LLM output is being seriously challenged however.

https://llmlitigation.com/case-updates.html

Personally I have roughly zero trust in US courts on this type of issue but we'll see how it goes. Arguably there are cases to be made where LLM:s cough up code cribbed from repos with certain licenses without crediting authors and so on. It's probably a matter of time until some aggressively litigious actors do serious, systematic attempts at getting money out of this, producing case law as a by product.

Edit: Oh right, Butterick et al went after Copilot and image generation too.

https://githubcopilotlitigation.com/case-updates.html

https://imagegeneratorlitigation.com/case-updates.html

This comes down to a question of what one can prove. NNs are necessary not explainable and none of this would have much evidence to show in court.

Sure there's evidence: Your statements about this when challenged. And perhaps to a degree the commit log, at least that can arouse suspicion.

Sure, you can say "I'd just lie about it". But I don't know how many people would just casually lie in court. I sure wouldn't. Ethics is one thing, it takes a lot of guts, considering the possible repercussions.

"I do not recall"

Yup, Gates style would work. But billionaires have a tendency to not get into serious trouble for lying to the public, a court, congress and what not. Commoners very much do.

How is it different from the cloud? Plenty startups store their code on github, run prod on aws, and keep all communications on gmail anyway. What's so different about LLMs?

I think it's a combination of a fundamental distrust of the model makers and a history of them training on user data with and without consent.

The main players all allow some form of zero data retention but I'm sure the more cautious CISO/CIOs flat out don't trust it.

I think that using something like Claude on Amazon Bedrock makes more sense than directly using Anthropic. Maybe I'm naive but I trust AWS more than Anthropic, OpenAI, or Google to not misuse data.

>How is it different from the cloud? Plenty startups store their code on github, run prod on aws, and keep all communications on gmail anyway. What's so different about LLMs?

Those plenty startups will also use Google, OpenAi or the built in Microsoft AI.

This is clearly for companies that need to keep the sensitive data under their control. I think they also get support with adding more training to the model to be personalized for your needs.

It’s not different. If you have a confidentiality requirements like that, you also don’t store your code off-premises. At least not without enforceable contracts about confidentiality with the service provider, approved by the client.

Most my clients have the same requirement. Given the code bases I see my competition generating, I suspect other vendors are simply violating this rule.

I have good results running Ollama locally with olen models like Gemma 3, Qwen 3, etc. The major drawback is slower inference speed. Commercial APIs like Google Gemini are so much faster.

Still, I find local models very much worth using after taking the time to set them up with Emacs, open-codex, etc.

I would take any such claim with a heavy rock of salt because the usefulness of AI is going to vary drastically with the sort of work you're tasked with producing.

You can set up your IDE to use local LLMs through e.g. Ollama if your computer is powerful enough to run a decent model.

Are your clients not on AWS/Azure/GCP? They all offer private LLMs out of the box now.

That was my question too.

Personally I am trying to see if we can leverage AI to help write design documents instead of code, based on a fairly large library of human (poorly) written design documents and bug reports.

I also kind of don't understand how it seems everyone is using AI for doing their homework. I haven't had a teacher yet which would have approved any AI usage.

Same process, less people being called out for "cheating" in a professional setting.

Have you tried using private inference that uses GPU confidential computing from Nvidia?

premises, not premise.

https://www.grammar-monster.com/easily_confused/premise_prem...

Also it’s like saying you can host a database on your Mac.

Unless you have experience hosting and maintaining models at scale and with an enterprise feature set, then I believe what they are offering is beyond (for now) what you’d be able put up on your own.

I have an M4 Mac Mini with 24GB of RAM. I loaded Studio.LM on it 2 days ago and had Mistral NeMo running in ten minutes. It's a great model, I need to figure out how to add my own writing to it, I want it to generate some starter letters for me. Impressive model.

I think many in this thread are underestimating the desire of VPs and CTOs to just offload the risk somewhere else. Quite a lot of companies handling sensitive data are already using various services in the cloud and it hasn't been a problem before - even in Europe with its GDPR laws. Just sign an NDA or whatever with OpenAI/Google/etc. and if any data gets leaked they are on the hook.

Good luck ever winning that one. How are you going to prove out a data leak with an AI model without deploying excessive amounts of legal spend?

You might be talking about small tech companies that have no other options.

> Btw, you can also run Mistral locally within the Docker model runner on a Mac.

Efficiently? I thought macOS does not have API so that Docker could use GPU.

I haven't/wouldn't use it because I have a decent K8S ollama/open-webui setup, but docker announced this a month ago: https://www.docker.com/blog/introducing-docker-model-runner

Hmm, I guess that is not actually running inside container/ there is no isolation. Some kind of new way that mixes llama.cpp , OCI format and docker CLI.

What's the point when we can run much powerful models now? Qwen3 , Deepseek

It would be short-termist for Americans or euros to use chinese-made models. Increasing their popularity has an indirect but significant cost in the long term. china "winning AI" should be an unacceptable outcome for America or europe by any means necessary.

Why would that be? I can see why Americans wouldn't want to do that, but Europeans? In the current political climate, where the US openly claims their desire to annex European territory and so on? I'd rather see them prefer a locally hostable open source solution like DeepSeek.

My two cents, as European, is that since we are more and more asking to LLMs for information, it wouldn't be wise to let a foreign country, not even truly democratic, to choose the information we get.

The Chinese don't get any of information if we use self-hosted DeepSeek or Qwen. They are open-source. You can run them in an air-gapped environment that can't phone home.

But their models are gimped by bad censoring. At least I can still ask chatgpt how many innocent civilians America has bombed.

There are plenty of other ways to run Mistral models on a Mac. I'm a big fan of Mistral Small 3.1.

I've run that using both Ollama (easiest) and MLX. Here are the Ollama models: https://ollama.com/library/mistral-small3.1/tags - the 15GB one works fine.

For MLX https://huggingface.co/mlx-community/Mistral-Small-3.1-24B-I... and https://huggingface.co/mlx-community/Mistral-Small-3.1-24B-I... should work, I use the 8bit one like this:

  llm install llm-mlx
  llm mlx download-model mlx-community/Mistral-Small-3.1-Text-24B-Instruct-2503-8bit -a mistral-small-3.1
  llm chat -m mistral-small-3.1

The Ollama one supports image inputs too:

  llm install llm-ollama
  ollama pull mistral-small3.1
  llm -m mistral-small3.1 'describe this image' \
    -a https://static.simonwillison.net/static/2025/Mpaboundrycdfw-1.png

Output here: https://gist.github.com/simonw/89005e8aa2daef82c53c2c2c62207...

Simon, can you recommend some small models that would be usable for coding on a standard M4 Mac Mini (only 16G ram) ?

That's pretty tough - the problem is that you need to have RAM left over to run actual applications!

Qwen 3 8B on MLX runs in just 5GB of RAM and can write basic code but I don't know if it would be good enough for anything interesting: https://simonwillison.net/2025/May/2/qwen3-8b/

Honestly though with that little memory I'd stick to running against hosted LLMs - Claude 3.7 Sonnet, Gemini 2.5 Pro, o4-mini are all cheap enough that it's hard to spend much money with them for most coding workflows.

How about on an MacBook Pro M2 Max with 64GB RAM? Any recommendations for local models for coding on that?

I tried to run some of the differently sized DeepSeek R1 locally when those had recently come out, but couldn’t manage at the time to run any of them. And I had to download a lot of data to try those. So if you know a specific size of DeepSeek R1 that will work on 64GB RAM on MacBook Pro M2 Max, or another great local LLM for coding on that, that would be super appreciated

I really like Mistral Small 3.1 (I have a 64GB M2 as well). Qwen 3 is worth trying in different sizes too.

I don't know if they'll be good enough for general coding tasks though - I've been spoiled by API access to Claude 3.7 Sonnet and o4-mini and Gemini 2.5 Pro.

How do you determine peak memory usage? Just look at activity monitor?

I've yet to find a good overview of how much memory each model needs for different context lengths (other than back of the envelope #weights * bits). LM Studio warns you if a model will likely not fit, but it's not very exact.

MLX reports peak memory usage at the end of the response. Otherwise I'll use Activity Monitor.

I'm also trusting `get_peak_memory` + some small buffer for now.

Still, it reports accurate peak memory usage for tensors living on GPU, but seems to miss some of the non-Metal overhead, however small (https://github.com/aukejw/mlx_transformers_benchmark/issues/...).

I imagine that this in quantized form would fit pretty well and be decent. (Qwen R1 32b[1] or Qwen 3 32b[2])

Specifically the `Q6_K` quant looks solid at ~27gb. That leaves enough headroom on your 64gb Macbook that you can actually load a decent amount of context. (It takes extra VRAM for every token of context you need)

Rough math, based on this[0] calculator is that it's around ~10gb per 32k tokens of context. And that doesn't seem to change based on using a different quant size -- you just have to have enough headroom.

So with 64gb:

- ~25gb for Q6 quant

- 10-20gb for context of 32-64k

That leaves you around 20gb for application memory and _probably_ enough context to actually be useful for larger coding tasks! (It just might be slow, but you can use a smaller quant to get more speed.)

I hope that helps!

0: https://huggingface.co/spaces/NyxKrage/LLM-Model-VRAM-Calcul...

1: https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-32...

2: https://huggingface.co/Qwen/Qwen3-32B-GGUF

There are plenty of smaller (quantized) models that fit well on your machine! On a M4 with 24GB it’s already possible to comfortably run 8B quantized models.

Im benchmarking runtime and memory usage for a few of them: https://aukejw.github.io/mlx_transformers_benchmark/

With around 4.6 GiB model size the new Qwen3-8B quantized to 4-bit should fit comfortably in 16 GiB of memory: https://huggingface.co/mlx-community/Qwen3-8B-4bit

16GB on a mac with unified memory is too small for good coding models. Anything on that machine is severely compromised. Maybe in ~1 year we will see better models that fit in ~8gb vram, but not yet.

Right now, for a coding LLM on a Mac, the standard is Qwen 3 32b, which runs great on any M1 mac with 32gb memory or better. Qwen 3 235b is better, but fewer people have 128gb memory.

Anything smaller than 32b, you start seeing a big drop off in quality. Qwen 3 14b Q4_K_M is probably your best option at 16gb memory, but it's significantly worse in quality than 32b.

What do you use to interface with Qwen?

I have LMStudio installed, and use Continue in VSCode, but it doesn't feel nearly as feature rich compared to using something like Cursor's IDE, or the GitHub Copilot plugin.

Continue can be your autocomplete provider – and use a smaller and faster model. Something like Cline (or Roo or Kilocode or another fork) would be the more Cursor-like assistant there.

Strange idea, but if I'd like to set up a solid LLM for use in my home network, how much processing power would I need for a multi-purpose model?

A Raspberry Pi? And old ThinkPad? A fully speced-out latest gen Macbook?

edit: One of those old Mac Pros?

That’s what I tried initially, an old black tin can Mac Pro, but it couldn’t do it. Next splashed on an m2 ultra 64gb mpro, runs ollama with qwen3 32b - reverse shell into the localhost with open web-ui and automatic111 and voila AI on my home network

Hm, that seems like a lot of power use. I thought I could get away with somewhat less.

Run Mistral 7b in under 4gb ram:

https://github.com/garagesteve1155/Overload

(As announced this morning in the FB group "Dull Men's Club!)

Why not use confidential computing based offerings like Azure's private inference for privacy concerns?

Game changer feels a bit strong. This is a new entry in a field that's already pretty crowded with open source tooling that's already available to anyone with the time and desire to wire it all up. It's likely that they execute this better than the community-run projects have so far and make it more approachable and Enterprise friendly, but just for reference I have most of the features that they've listed here already set up on my desktop at home with Ollama, Open WebUI, and a collection of small hand-rolled apps that plug into them. I can't run very big models on mine, obviously, but if I were an Enterprise I would.

The key thing they'd need to nail to make this better than what's already out there is the integrations. If they can make it seamless to integrate with all the key third-party enterprise systems then they'll have something strong here, otherwise it's not obvious how much they're adding over Open WebUI, LibreChat, and the other self-hosted AI agent tooling that's already available.

> crowded with open source tooling that's already available to anyone with the time and desire to wire it all up.

Those who don't have the time and desire to wire it all up probably make up a larger part of the market than those who do. It's a long-tail proposition, and that might be a problem.

> I have most of the features that they've listed here already set up on my desktop at home

I think your boss and your boss' boss are the audience they are going for. In my org there's concern over the democratization of locally run LLMs and the loss of data control that comes with it.

Mistral's product would allow IT or Ops or whatever department to set guardrails for the organization. The selling point that it's turn-key means that a small organization doesn't have to invest a ton of time into all the tooling needed to run it and maintain it.

Edit: I just re-read your comment and I do have to agree though. "game-changer" is a bit strong of a word.

Actually you shouldn't be running LLMs in Docker on Mac because it doesn't have GPU support. So the larger models will be extremely slow if they'll even produce a single token.

I think the the standard setup for vscode continue for ollama is already 99% of ai coding support I need. I think it is even better than commercial offerings like cursor, at least in the projects and languages I use and have tested it.

We had a Mac Studio here nobody was using and it we now use it as a tiny AI station. If we like, we could even embed our codebases, but it wasn't necessary yet. Otherwise it should be easy to just buy a decent consumer PC with a stronger GPU, but performance isn't too bad even for autocomplete.

Which models are you using?

How many is many? Literally all of them use cloud services.

I really don't see the big deal. Gemini also allows on-prem in similar fashion: https://cloud.google.com/blog/products/ai-machine-learning/r...