Hey everyone, happy Friday evening.

Yesterday a French lab handed over the files for a model with a trillion parameters, and then quietly kept one string attached.

The model is Mistral Large 4, nicknamed Le Chonk, and from October 29 anyone can download its weights and run it on their own servers. Google announced its newest model two days earlier and handed it to vetted cyber defenders first.

Quick take: Mistral Large 4 is now open weight, a 1 trillion-parameter model with a 1 million-token context window, at half its preview API price. You can use it, change it, and build products on it, but selling access to the model itself needs a separate license. For most people the API is still the sane way in.

Mistral: Mistral Large 4 — 1 trillion parameters (49 billion active), 1 million-token context window, $0.70 / $2.10 API price per million tokens. Download the weights, or use the API.

MiniMax: M3.1 Flash — 428 billion parameters (23 billion active), 1 million-token context window, $0.15 / $0.60 API price per million tokens. Download the weights, or use the API.

Gemini: 4 Carbon — size and context not published, $2 / $10 intro API pricing per million tokens, gated to cyber defenders first.

What actually shipped

Mistral first showed Large 4 in a public preview on October 6, API only, with the weights promised by the end of October. The weights landed on October 29, and they are not just a re-upload of the preview. Mistral says this is an improved checkpoint after three more weeks of extra training.

The shape of it is a mixture-of-experts design. Only a slice works on each token, 49 billion parameters active at a time, 52 billion if you count the embedding and output layers.

It takes text and images in and writes text out, and it can hold up to 1 million tokens of context in one request.

On Mistral's own numbers, it leads Cognition's FrontierCode 2.0 at 19.7 percent, and it scores 48 percent on the Artificial Analysis Coding Agent Index while using 47 percent fewer tokens than the rival tied with it.

On cyber tests it scores 51 percent on the Artificial Analysis Cyber Index, and Mistral says it refused 96.6 percent of malicious cyber requests in its tests.

I will say this plainly, because it matters. Those are Mistral's numbers, not independent checks. Nobody outside Mistral has had time to test the released weights properly yet, and the full technical report is still coming.

The price cut matters more than the size

A trillion parameters sounds huge, and it is, but the number that changes real decisions is the API price. At preview, Large 4 cost $1.36 per million input tokens and $4.18 per million output tokens. Now it is $0.70 and $2.10, with cached input at $0.07. That is roughly half, overnight.

Compare that with the two models I wrote about last week. MiniMax M3.1 Flash is far cheaper through its API at $0.15 and $0.60, and its weights have been out since October 23. Gemini 4 Carbon is $2 and $10 on intro pricing, and still gated to defenders.

Mistral lands in the middle on price and at the top on openness.

The one condition in the small print

The weights ship under the Mistral Ecosystem Open-Weight License, a modified MIT license. You can use the model, change it, fine-tune it, and sell products built on it.

The condition kicks in if your business is selling access to the model itself as a hosted service. In that case you need a separate commercial license from Mistral first. Building a product that uses the model inside one feature is not the same thing, and internal use that keeps the model away from third parties is not caught either.

But running a lookalike API and charging for it is. Mistral is giving away a flagship model and keeping the one revenue stream that would let a rival clone its API business for nothing. Just do not mistake "open weight" for "no rules at all".

Why Europe cares, and why the cyber angle cuts both ways

Mistral keeps pitching this as a sovereignty story, and I think that pitch lands this time. The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers.

For a team that does not want its work routed through a US or Chinese provider, a model they can run on their own servers is a real answer. Nobody can remotely switch it off or quietly change the version under you.

The cyber angle is where it gets two-sided. The company pitches Large 4 for defensive security work, finding flaws and fixing code. The same download, though, is open to everyone, including people who want it for attacks.

Mistral's answer is that refusal rate, 96.6 percent of malicious requests turned down in its tests. I am not fully sold yet, until independent testers run the released weights.

Bottom line

If you are a developer, try it through the API first. At $0.70 and $2.10 it is cheap enough to test against whatever you use now, and you skip the hardware headache.

If you run a team that cares about control, data staying in-house, or a model nobody can take away, the download is the real prize, just budget for it properly. The setup examples talk about a server with eight GPUs.

The open-weight wave this month is now hard to ignore. MiniMax on October 23, Mistral on October 29, and Kimi K3.1 Pro's weights due by November 4. I will be watching the independent scores when they land.

What would you do with it first, run it through the API, or try to host it yourself? And does the license condition change your answer?

Cover image: AI-generated

Originally published on PeakD: https://peakd.com/hive-167922/@worthmind/mistral-just-gave-away-its-biggest-model-and-the-small-print-is-the-interesting-part