Daily Issue #6 · 7 min read

Mistral just gave away a trillion parameter model. Here's who should care.

Mistral Large 4 is now free to download. It is not the smartest open model, but it may be the most efficient, and that changes who should care.

The short answer

What did Mistral release? On October 29, Mistral published the open weights of Mistral Large 4, nicknamed Le Chonk. It is a mixture of experts model with about 1 trillion total parameters, roughly 49 billion active per token, a 1 million token context window, and text plus image input. Anyone can download it from Hugging Face, fine tune it and use it commercially, unless they want to sell it as a hosted model service, which needs a separate license from Mistral. The independent firm Artificial Analysis ranks it as the top open weight model built outside China, but six points behind the Chinese leaders. My view: the score is not the story. The story is cost per task, and the fact that a serious Western option now exists for teams that must run AI on their own hardware.

What Mistral actually shipped

Mistral first showed Large 4 as a public preview earlier in October. The version released this week is an improved one, after what Mistral calls "just three weeks of additional reinforcement learning." According to the Hugging Face model card, the key facts are:

• About 1 trillion total parameters, with 128 routed experts and 4 active for each token, plus one shared expert.
• Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe.
• Three reasoning levels (low, medium and high), set per request.
• Official FP8 and NVFP4 checkpoints, with support from vLLM, SGLang, Transformers and llama.cpp on day one.
• Fluency in more than 160 languages, including every official language of the European Union.

The hosted API price is now $0.70 per million input tokens and $2.10 per million output tokens. For comparison, Google's Gemini 4 Carbon launched this week at an introductory $2 and $10.

Mistral's numbers and the independent numbers

Mistral says Large 4 is the best open weight model for coding. It reports 19.7% on Cognition's FrontierCode 2.0 and 41% on Terminal Bench 4.0. Artificial Analysis, which ran its own tests, reported 15% and 25% on those same two benchmarks. Neither side has explained the gap, so I treat them as two different setups rather than one side being wrong. But it is a good reminder that launch charts come from the company that wants you to buy.

On the Artificial Analysis Intelligence Index v5, Large 4 scores 26 at its high reasoning setting. That narrowly puts it first among open weight models developed outside China. DeepSeek V4.1 Flash and GLM 5.3 lead the open field six points higher. The top closed models sit far above all of them, with Gemini 4 Carbon at 59.

Read only the leaderboard and this looks like a middling result. I think that misses what changed.

The real gain is efficiency

Here is the part I find most useful. Artificial Analysis says Large 4 uses about 82,000 output tokens per Intelligence Index task, while the seven open models that score higher use between 113,000 and 225,000. VentureBeat, reporting the same evaluation, says the cost to run each index task fell to $2.06 from $5.54 for the preview, a 63% cut. Artificial Analysis also reports its hallucination rate on the AA Omniscience test fell to 28%, from 41% for the preview.

Fewer tokens means lower bills and faster answers, and that matters most for agents. An agent that thinks twice as long on every step costs twice as much, and the difference compounds across a long run. A model that lands a few points lower but finishes the job with half the tokens can be the better business choice for routine work. One fair caveat, raised by commenters on X: each model was tested in its own coding harness, so part of the token gap may come from the harness, not the model.

Open, but read the license

The Mistral Ecosystem Open Weight License is a modified MIT license. You can use, modify, fine tune and sell products built on the model. The one condition is that a business offering the model as a hosted service needs a commercial license first. The two biggest threads in the Hacker News discussion argued that this makes it "weights available" rather than truly open.

For most teams the label matters less than the terms. If you want to run it inside your own product or your own cloud, the license is generous. If you are a cloud or inference provider, it is not free to resell, and you should talk to Mistral before building on it.

Running it is still a big job

"Downloadable" does not mean "runs on your laptop." Unsloth cofounder Daniel Han put the FP8 weights at 1.1 terabytes, and the serving examples on the model card assume a node of 8 GPUs. Unsloth's 1 bit version shrinks that to 223 gigabytes, which Unsloth says can run on a Mac Studio while keeping 91% top 1 accuracy. For almost everyone else, the realistic paths are the Mistral API, OpenRouter, Microsoft Foundry or NVIDIA hosted endpoints, all of which Mistral lists.

The security tension

Mistral pitches Large 4 as strong at defensive security work. It reports 51% on the Artificial Analysis Cyber Index and says the model refuses 96.6% of malicious cyber requests on standard test sets. It also says something unusually candid: after watching live traffic during the preview, "we see a potential for uplift to conduct real attacks," and it is changing its guardrails.

That tension is real. Google is rationing Gemini 4 Carbon to vetted cyber defenders because of the same risk. Mistral has handed out weights that anyone can fine tune, and guardrails trained into open weights can be weakened by whoever holds them. I do not think that makes open release wrong, since defenders also gain from models they can audit and run privately. But it means the safety of an open model depends on the people deploying it, not only on the lab that trained it.

Who should pay attention

• Teams in regulated sectors such as banking, health and government that cannot send data to an outside API. This is now a credible model they can host themselves.
• Anyone running coding or browsing agents at high volume, where token efficiency drives the bill.
• Builders who want a fallback that no single provider can change or switch off.

If you happily use a closed API today and your work needs the very best reasoning, this release does not change your default. Test it as a cheaper option for routine steps, not as a replacement for your top model.

My takeaway

Mistral Large 4 is not the smartest model you can use, and Mistral's own benchmark claims run ahead of the independent ones. But it is a trillion parameter model that you can own, customize and run under your own rules, at a fraction of the token cost of models that score higher. For a growing number of teams, control and cost per task matter more than a leaderboard rank. That is the gap Mistral is filling, and I expect more companies to test it seriously for exactly that reason.

Quick notes

• OpenAI rewrites its Model Spec around five values: OpenAI updated its Model Spec on October 29, the public document describing how its models should behave. It is now organized around faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI says the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance.
• Moonshot launches Kimi K3.1 Pro with weights promised next week: Moonshot AI launched Kimi K3.1 Pro on October 30, describing it as its most capable 3 trillion parameter model to date, with a 1 million token context. A Highspeed version offers what Moonshot calls roughly 23 times faster output. Both are available through Kimi's apps and API now, and Moonshot says full weights will be released by November 4.

Common questions

What is Mistral Large 4?

Mistral Large 4, nicknamed Le Chonk, is Mistral's largest model. It is a mixture of experts model with about 1 trillion total parameters and roughly 49 billion active per token, a 1 million token context window, and text and image input. Mistral released its open weights on October 29, 2026.

Is Mistral Large 4 open source?

Its weights are free to download under the Mistral Ecosystem Open Weight License, a modified MIT license. You can use, modify and commercialize it, but a business that offers the model as a hosted service needs a separate commercial license from Mistral. Some developers argue this makes it weights available rather than open source.

How much does Mistral Large 4 cost?

The weights are free. On the Mistral API it costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

How does Mistral Large 4 compare with other open models?

Artificial Analysis scores it 26 on its Intelligence Index v5, the highest among open weight models built outside China. DeepSeek V4.1 Flash and GLM 5.3 score six points higher. Artificial Analysis also found it uses far fewer tokens per task than the higher scoring open models.

Can I run Mistral Large 4 on my own computer?

Not easily. The FP8 weights are about 1.1 terabytes and the official serving examples use a node of 8 GPUs. Unsloth offers a 1 bit version of about 223 gigabytes that Unsloth says can run on a Mac Studio. Most people will use a hosted API instead.