Mistral ships the weights of Le Chonk, with an asterisk in the licence
Mistral published the open weights of Mistral Large 4 on 29 October under a modified MIT licence. The coding scores moved; the intelligence ranking did not.
Vlad Makarovreviewed and published
Mistral published the weights of Mistral Large 4 on 29 October, three weeks after the 6 October public preview that introduced the model as its largest yet. The post is titled We release the weights of Le Chonk, and it reads as a follow-up rather than a launch: the same 1-trillion-parameter, natively multimodal mixture of experts, now with three reasoning modes and an actual licence attached. What moved is the coding column, and the price. What did not move, at least in public, is the aggregate score that made the preview eighth among open-weight models.
What actually landed on Hugging Face
The repository is ungated and was last written at 14:51 UTC on 29 October, though created on 30 September, three weeks before the preview. It holds 54 safetensors shards totalling about 1.06 TB, with weights in fp8 and the vision tower and output head outside the quantisation. The architecture string is MistralLarge4ForConditionalGeneration and the pipeline tag image-text-to-text: the multimodal claim is in the file layout, not just the copy. The model card claims 160+ languages.
Two mirrors appeared on day one: an NVFP4 variant from Mistral, a GGUF build from Unsloth. Neither makes it local. A commenter on Hacker News put the practical version plainly: "I have no idea how you could shrink this down to fit on any type of prosumer hardware though." The download button is open; the addressable hardware is still a rack.
The licence is MIT with a services carve-out
Mistral calls the Mistral Ecosystem Open-Weight license a modified MIT licence, and mostly it is: use, modify, distribute, sublicense, sell, further train, with one condition on top. A licensee, or its affiliates, that operates a Model as a Service business must sign a separate commercial licence with Mistral before commercial use of the model or a derivative. The file defines that as giving a third party access for inference or post-training in a way that lets them control inputs, parameters or training data, carving out products that embed model capabilities inside specific features.
That is narrower than "open weights" implies: it is a field-of-use restriction, and the Open Source Definition does not permit those. Hugging Face's metadata agrees — the repo's licence tag reads "other", not "mit". A commenter made the same point without the legal vocabulary: "It's not open weight, it's weights available as they use a modified MIT license that is not OSI."
The carve-out is not hidden; it sits in paragraph two of a short file. But the gap between downloading a model and selling inference is the one the preview left open, when the licence had no name.
Coding: the numbers moved on Cognition's benchmark
By Mistral's account, three weeks of extra reinforcement learning produced these gains. On Cognition's private FrontierCode 2.0, which asks whether a repository maintainer would merge an agent's change, Mistral reports:
| Model | FrontierCode 2.0 |
|---|---|
| Mistral Large 4 | 19.7% |
| DeepSeek V4.1 Flash | 19.3% |
| GLM-5.3 | 17.8% |
| Kimi K3 | 17.3% |
| GLM-5.3 Flash | 14.2% |
A 0.4-point lead on a private benchmark, run by a firm that also sells coding agents, is thin ground, and nobody outside Cognition can rerun it. The second number is more interesting: on the Artificial Analysis Coding Agent Index Mistral reports 48%, "effectively tied" with GLM-5.3 in the post's own wording, while using 47% fewer tokens per task. Terminal-Bench 4.0 rises 13 points over the preview to 41%, with 30% fewer output tokens. Efficiency matters more than ranking: verbosity was the specific complaint against the preview, which burned about 200 million output tokens on the Intelligence Index against a class median near 81 million.
Surge Code's blind pairwise runs point the same way, with the usual caveat about who paid: 64% of prompts preferred over GLM-5.3 and Kimi K3, 56% over DeepSeek V4.1 Flash, 22% over Claude Opus 5.5. A vendor-commissioned panel: directional, not a verdict.
Best for coding is not the same as most intelligent
The headline claim is that Mistral Large 4 is the best open-weight model for coding. That may hold. It is a different claim from being the most capable open model overall, and the post offers no new Artificial Analysis Intelligence Index score to support the second. The preview scored 38.4, eighth among open-weight models and behind seven Chinese ones, at $1.13 per task against $0.27 for a model that scored higher. Nothing published on 29 October revises that.
Pricing did change, and sharply: $0.70 per million input tokens, $0.07 cached input and $2.10 output, against $1.36, $0.14 and $4.18 at preview. That is roughly half across the board, and with the token-efficiency claims it answers the cost-per-task critique rather than rebutting it. Whether that lowers cost per completed task depends on how the modes are used; no new per-task figure has been published.
Cyber: guardrails up, caveat intact
Mistral reports 51% on the Artificial Analysis Cyber Index, and the comparison bars beside it run from 34% to 57%. That is mid-pack: clear of Beam 1.0, GLM-5.3 and Kimi K3, level with GLM-5.3 Flash, behind MiMo-V2.6-Pro and DeepSeek V4.1 Flash. It claims the lead on refusals — an average 96.6% refusal rate on malicious cyber prompts across JailbreakBench, StrongREJECT and AgentHarm, and 97.1% attack resistance on the B3 agent-security benchmark, up from 93.3%.
One sentence matters more than any chart above it. Mistral says that while monitoring live traffic it saw the model refuse black-box attacks against live targets, but that it does "see a potential for uplift to conduct real attacks", and that it is modifying guardrails accordingly. A company releasing a trillion-parameter security model, and saying in the same document that it could help someone attack a real system, is being honest about a risk it cannot price. The specialised cyber variant goes only to vetted design partners under contract, restricted to systems they own or are authorised to test.
What the release settles, and what it does not
The weights shipping matters for one reason the marketing does not stress: the model can now be examined. Artificial Analysis classified the preview as proprietary; that objection is gone. Independent reruns at matched effort are now possible, and the reasoning-mode table gives evaluators a knob that did not exist three weeks ago: low, medium, high, high by default. The post publishes no technical report, so the data mix and architecture choices remain claims, not documentation.
Reception was large and unsurprised: 1,048 points and 305 comments on Hacker News, much of it about quantisation, memory and whether modified MIT counts as open source. That is the right argument. Mistral has shipped a strong coding model, priced it to compete, and published terms that let almost anyone use it except the group that most wants to — companies whose business is serving models to other people. Europe has an open-weight model worth arguing about, which is progress. It does not have one that leads on general capability, and this release does not say otherwise.
