Mistral Large 4 "Le Chonk" Weights Are Out: What Changed Since the Preview, and What Is Still Claimed

Mistral has published the open weights of Mistral Large 4, the 1-trillion-parameter model it nicknamed Le Chonk, three weeks after the API preview that we covered in our preview post. In its announcement, the lab says additional reinforcement learning since the preview made the model the best open-weight option for coding and for defensive cybersecurity, and that it now uses fewer tokens to get there. Hugging Face published a launch guide the same day, with checkpoints and inference recipes.

The catch worth knowing up front: almost every performance number comes from Mistral, and the lab's post and the Hugging Face post do not agree on several of them. This article separates what is announced, what is verifiable today, and what you can do with the weights.

Update (October 29, 2026): Artificial Analysis puts Large 4 at 26, behind Chinese open models

According to The Decoder's summary of Artificial Analysis data, Mistral Large 4 scores 26 on the Intelligence Index (up one point from the 25-point preview). That beats the median of 23 for similarly sized open models but trails DeepSeek V4.1 Flash and GLM-5.3 (32 each) and MiMo-V2.6-Pro (30). Mistral is the top open-weight model from outside China on that ranking, not the top open model overall.

• Coding: Terminal-Bench 4.0 at 41% (GLM-5.3: 40%); Coding Agent Index 48%, nearly tied with GLM-5.3 while using about 47% fewer tokens.
• Cost and speed: about $2.06 per index task (preview: $5.54), 161 output tokens per second, and API prices of $0.70 input and $2.10 output per million tokens. DeepSeek V4.1 Flash is far cheaper per task at $0.60.
• Human preference: in Surge's blind tests, Large 4's code beat GLM-5.3, Kimi K3 and DeepSeek, but testers chose it over Claude Opus 5.5 only 22% of the time.

Note the small mismatch with the Hugging Face figure quoted below (a 49.8% agent index): different snapshots of the same leaderboard, so check the live Artificial Analysis page. See also our Intelligence Index v5 explainer.

TL;DR: the questions people ask first

What changed between the preview and the weights?

Mistral Large 4 entered public preview earlier this month with an API and a promise of open weights by the end of October. The new post says three more weeks of reinforcement learning (RL) produced "massive improvements" in coding and defensive security, and that "gains continue to accelerate, with no signs of saturation." Mistral also replaced the old "None" reasoning option with a "Low" mode, giving three settings: Low for simple tasks, Medium for everyday work and High (the default) for hard problems.

That means the released model is not the same checkpoint as the preview. If you tested the preview through the API and drew conclusions, retest, because Mistral says reasoning efficiency changed, meaning fewer output tokens for similar quality.

The benchmark claims, with sources

Everything in this table is reported by Mistral in its own post unless marked otherwise.

Three of these come from third-party evaluators (Cognition, Artificial Analysis, Surge), but Mistral chose which results to show and how to frame them. FrontierCode 2.0 is a private benchmark, so you cannot rerun it. The Artificial Analysis numbers can be checked on that site's own pages, and we explained how that index works in our Intelligence Index v5 post.

Where the two official posts disagree

The Hugging Face guide, written by the Hugging Face team rather than Mistral, lists different numbers for several of the same tests:

• Terminal-Bench 4: Hugging Face lists 28.3%. Mistral's new post says 41%, a 13-point gain. Our preview post recorded 28.3% for the preview model, so the Hugging Face figure looks like the earlier checkpoint.
• B3 attack resistance: Hugging Face says 93.3%; Mistral says 97.1% now, versus 93.3% in preview. Again the Hugging Face figure matches the preview.
• Cyber: Mistral's post gives an Artificial Analysis Cyber Index of 51%. Hugging Face gives an 82% score on one sub-task, vulnerability reproduction and patching, which it calls the highest reported within the index. These can both be true, because one is the overall index and the other a component, but a reader skimming either post alone would come away with very different impressions.
• Active parameters: Mistral and Hugging Face both say about 49B active, while the Hugging Face code sample uses the model id Mistral-Large-4-1T-A52B, which implies 52B. We could not tell from the posts which is exact; the model card should settle it.

None of this suggests bad faith. It is a day-one launch with a lot of parties publishing at once. But it is a good reason to wait for the model card and independent runs before quoting a single number as settled.

Why the cyber angle is the real story

Most open-weight launches compete on coding and math. Mistral made defensive security the headline, and the argument is worth understanding. Reproducing a vulnerability before patching it looks, to a safety filter, a lot like attacking. Hugging Face notes that closed models often refuse such requests, and that an API-dependent team can lose access mid-incident. With open weights, the organization sets its own usage policy and runs the model on its own infrastructure.

Mistral quotes security vendors including CrowdStrike, Cisco, Cloudflare, Elastic, Fortinet, Orange Cyberdefense, Rubrik, Zscaler and Booz Allen, plus the EU's ENISA, which is quoted as experimenting with both closed and open-weight models. These are vendor endorsements published on Mistral's page, not independent audits.

Mistral is also candid about the downside. After watching live preview traffic it says it saw "a potential for uplift to conduct real attacks" and is modifying its guardrails. It is giving vetted design partners such as Orange Cyberdefense access to a specialized cyber variant under contractual limits. Open weights, however, cannot be recalled: anyone can fine-tune away built-in refusals. That is the central tension of releasing a strong cyber model openly, and it echoes the debate around Google's Gemini 4 Carbon cyber launch, which went the opposite way, with access limited to defenders.

Architecture in plain language

Per the Hugging Face guide, the text decoder has 72 layers and 128 experts per mixture-of-experts layer, with 4 routed experts plus one shared expert active per token, so about 5 experts fire for each token. It pairs Multi-head Latent Attention, which compresses the key-value cache, with a DeepSeek-style sparse attention indexer that picks up to 2,048 earlier positions per query. A 1.6B-parameter Pixtral vision encoder handles images. The practical effect is that only a small slice of the 1T parameters runs per token, which cuts compute but not memory: all weights must still be loaded.

How to run it

An open box releasing seeds toward waiting hands, picturing the Mistral Large 4 open weights download

Hugging Face says Large 4 has day-zero support in Transformers, vLLM and llama.cpp, with serverless access through Inference Providers and dedicated Inference Endpoints. The vLLM recipe for the NVFP4 checkpoint uses these key flags:

vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 \
  --config-format mistral --load-format mistral \
  --tensor-parallel-size 8 --enable-expert-parallel \
  --max-model-len 131072

Notes from the guide: the NVFP4 checkpoint needs NVIDIA Blackwell GPUs, the model ships with a drafter for speculative decoding speedups in vLLM, and the Transformers path needs a recent transformers release and torchrun across eight processes. If you do not own that hardware, the hosted Inference Providers route or the Mistral API is the realistic path. Our preview post has a section on why a 1T open model is not the same as run-it-yourself.

How does it compare with other open models?

Hugging Face says that as of October 29, 2026 the model is the fifth best open-weight model on the Artificial Analysis Intelligence Index, and that its Coding Agent Index of 49.8% is ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. Mistral's own framing is narrower: top among open models for coding, tied with GLM-5.3 on the agent index, and ranked behind only Claude Opus 5.5 in Surge's human-preference test.

That combination, fifth overall but first or tied on coding agents, is plausible for a model tuned with RL on agentic tasks, but it also shows why one headline cannot capture it. For the geopolitics of that race, including Chinese labs' position and distillation claims, see our DeepSeek, Moonshot and GPT-6 distillation fact check. For another case of a smaller open model beating much larger ones on a narrow task, see Samaya Sterling vs Fable 5.1.

What this means for what you build or pay

• Coding agents: if token cost dominates your bill, the claim of 47% fewer tokens at similar index score matters more than raw accuracy. Test on your own repository with all three reasoning modes, because Low may be enough for routine edits.
• Security teams: a self-hosted model avoids sending code or logs to a third party and avoids refusals on legitimate triage. Pair any agent with a policy layer; see our look at Hugging Face SafeAgents and the AgentBeam agent security platform from the explainx.ai team for blocking dangerous tool calls before they run.
• Regulated and European buyers: the pitch is lineage. Mistral says the model was trained from scratch on European infrastructure, which matters for sovereignty audits. Read the license, especially the model-as-a-service limits.
• Anyone shipping a hosted product: the license restriction on offering the model as a service may bite. Get it reviewed before you build a resale API on it.

What is still unverified

1. Independent reproduction of the post-preview gains, especially Terminal-Bench 4.0 at 41% and the 51% Cyber Index.
2. The exact active parameter count and which checkpoint each Hugging Face number refers to.
3. The real-world abuse risk once guardrails can be fine-tuned away.
4. Whether API pricing stays put: Mistral's post lists $0.70 per million input tokens, $0.07 cached input and $2.10 output, but our preview post recorded launch and promotional prices, so check Mistral's pricing page.

Bottom line

The release itself is real and significant: a trillion-parameter, multimodal, openly licensed model from a European lab, with day-zero tooling support. The performance story is plausible but vendor-reported, and the two launch posts already disagree on some figures. Download it if you have the hardware or a provider, test it on your own tasks, and treat leaderboard-style claims as a starting point. We will add an update here when independent numbers or the full model card settle the open questions.

Related reading

• Mistral Large 4 preview: 1T parameters, 49B active
• Mistral and NVIDIA: open benchmarks for CAE and EDA agents
• Artificial Analysis Intelligence Index v5 and the open-weight gap
• Cognition FrontierCode 2.0: claimed vs verified
• Gemini 4 Carbon: a cyber model for defenders first
• DOJ questions on Nvidia and Hugging Face
• Official: Mistral announcement and Hugging Face launch guide

Specs, benchmarks and licensing are accurate as of October 29, 2026 and may change as Mistral and independent evaluators publish more.