Mistral ships Le Chat – enterprise AI assistant that can run on prem

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.

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.

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

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?

Run Mistral 7b in under 4gb ram:

https://github.com/garagesteve1155/Overload

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

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.)

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 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.

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.

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.

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.

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 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.

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.

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.

Black Forest Labs are the makers of FLUX, which for a while was the best open image model available (and generally a pretty strong image model). That said, now with a wave of Chinese models and the advent of autoregressive image models, I'm not sure how much that will stay true.

Agents are specialized via prompts and MCP now, and more and more rarely by model

That’s a perfectly valid idea in theory, but in practice you’ll run into a few painful trade-offs, especially in multi-user environments. Trust me, I'm currently doing exactly that in our fairly limited exploration of how we can leverage local LLMs at work (SME).

Unless you have sufficient VRAM to keep all potential specialized models loaded simultaneously (which negates some of the "lightweight" benefit for the overall system), you'll be forced into model swapping. Constantly loading and unloading models to and from VRAM is a notoriously slow process.

If you have concurrent users with diverse needs (e.g., a developer requiring code generation and a marketing team member needing creative text), the system would have to swap models in and out if they can't co-exist in VRAM. This drastically increases latency before the selected model even begins processing the actual request.

The latency from model swapping directly translates to a poor user experience. Users, especially in an enterprise context, are unlikely to tolerate waiting for a minute or more just for the system to decide which model to use and then load it. This can quickly lead to dissatisfaction and abandonment.

This external routing mechanism is, in essence, an attempt to implement a sort of Mixture-of-Experts (MoE) architecture manually and at a much coarser grain. True MoE models (like the recently released Qwen3-30B-A3B, for instance) are designed from the ground up to handle this routing internally, often with shared parameter components and highly optimized switching mechanisms that minimize latency and resource contention.

To mitigate the latency from swapping, you'd be pressured to provision significantly more GPU resources (more cards, more VRAM) to keep a larger pool of specialized models active. This increases costs and complexity, potentially outweighing the benefits of specialization if a sufficiently capable generalist model (or a true MoE) could handle the workload with fewer resources. And a lot of those additional resources would likely sit idle for most of the time, too.

Good idea. Then you could place another lighter-weight model in front of THAT, to figure out which model to use in order to find out which model to use.

It,'s LLMs, all the way down.

My guess is that this is basically what AI providers are slowly moving to. And this is what models seem to be doing underneath the surface as well now with Mixture of Experts (MoE).

I mean, the general purpose models already do this in a way, routing to a selected expert. It's a pretty fundamental concept for ensemble learning, which is what MOE experts are, effectively.

I don't see any reason you couldn't stack more layers of routing in front, to select the model. However, this starts to seem inefficient.

I think the optimal solution will eventually be companies training and publishing hyper-focused expert models, that are designed to be used with other models and a router. Then interface vendors can purchase different experts and assemble the models themselves, like how a phone manufacter purchases parts from many suppliers, even their compeditors, in order to create the best final product. The bigger players (e.g. Apple for this analogy) might make more parts in house, but even the latest iPhone still has Samsung chips in it in teardowns.

Yeah, I mean paid API access. You put a credit card in, and it's peanuts at the end of the month. Sorry I didn't specify. Good reminder that with free services you are the product!

Presumably one throat to choke logic applies here, particularly in Europe.

Explain more please? Is that a big phone/tiny laptop with long GPU connector? Is that a tiny model?

The big picture incongruence in the thread is using terms as patriotic, allegories to US states, which imo is but a US-centric projection. Even proponents don’t think of the EU to be a supreme government with federated states, and they certainly don’t think of ”Europeans” as a unified demographic. At best, the EU protects against stupid shit from other EU countries (tariffs, freedom of movement) and stupid shit against the outside, such as bullying by superpowers like Russia, China and recently also the US, or extremely large corporations who can take on smaller nation states.

The EU is more similar to NAFTA or five eyes, and culturally the loyalty is more similar to the US vs the anglosphere, like how Americans think of Australia, UK and Canada. Well, again, until recently. Things are changing fast.

>You probably want to replace Llama with Qwen in there. And Gemma is not even close.

Have you tried the latest, gemma3? I've been pretty impressed with it. Altho I do agree that qwen3 quickly overshadowed it, it seems too soon to dismiss it altogether. EG, the 3~4b and smaller versions of gemma seem to freak out way less frequently than similar param size qwen versions, tho I haven't been able to rule out quant and other factors in this just yet.

It's very difficult to fault anyone for not keeping up with the latest SOTA in this space. The fact we have several options that anyone can serviceably run, even on mobile, is just incredible.

Anyway, i agree that Mistral is worth keeping an eye on. They played a huge part in pushing the other players toward open weights and proving smaller models can have a place at the table. While I personally can't get that excited about a closed model, it's definitely nice to see they haven't tapped out.

[flagged]

[flagged]

Instead it disobeys commands, uses up your resources then you find it never belonged to you in the first place.

> In this case, the law has no moral valence

That's not how laws work.

There is a difference if you upload your data or your customers data.

There are countries in the EU where you get sued for less