Mistral Large 4 Is Now Open-Weight: What Changed, What the Benchmarks Say, and Who Should Care
Mistral has released the weights of Mistral Large 4, the 1-trillion-parameter, natively multimodal model it nicknames "Le Chonk." The model entered public preview earlier in October. The open-weight release is an updated version: Mistral says it spent about three more weeks on reinforcement learning after the preview, with a focus on coding and defensive cybersecurity.
The short version
- What it is: a 1T-parameter multimodal flagship model, trained from scratch on European infrastructure (per Mistral).
- What's new vs. the preview: better coding and security scores, plus better "reasoning efficiency", meaning fewer tokens per task.
- Where to use it: Mistral Studio, the Mistral API, OpenRouter, or as self-hosted weights from Hugging Face (served with vLLM or SGLang).
- License: a modified MIT license with one condition aimed at Model-as-a-Service companies (see our license explainer).
The numbers Mistral reports
All figures below are Mistral's own claims from its announcement. Independent replication isn't available yet, and Mistral says a full technical report is "coming soon."
| Benchmark | Reported result |
|---|---|
| Cognition FrontierCode 2.0 (private) | 19.7%, the top score in Mistral's chart, ahead of GLM-5.3 (19.3%) |
| Artificial Analysis Coding Agent Index | 48%, "effectively tied" with GLM-5.3 while using 47% fewer tokens per task |
| Terminal-Bench 4.0 | 41%, up 13 points from the preview, with 30% fewer output tokens |
| Artificial Analysis Cyber Index | 51% |
| BrowseComp | 84.5% (up from 75% at an early SFT checkpoint) |
| Finch (finance workflows) | 73.3%, second among open models to Kimi K3 |
| B3 Agent Security Benchmark | 97.1% of attacks resisted (93.3% in preview) |
On human preference, Mistral cites Surge's blind pairwise coding tests. In those, Large 4 beat GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked "behind only Claude Opus 5.5" among the models tested.
Three reasoning modes
Mistral says the model now has three reasoning modes, so you can trade cost against quality. The lightest mode handles simple tasks and replaces the old "None" mode. The heaviest is for "hardest use cases requiring high precision." If you pay per token, this matters: pick the lightest mode that passes your evals.
The cybersecurity angle (and its caveat)
Most of the announcement covers defensive security. It includes quotes from Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA and others. Mistral also openly says that while watching preview traffic it saw "a potential for uplift to conduct real attacks," and that it is tightening guardrails. A specialized pentest variant is going only to vetted partners under contract.
Should you switch?
- Self-hosters and regulated teams: this is the main audience. You get open weights, a European training lineage and a permissive license.
- Hobbyists: a 1T-parameter model needs data-center hardware to run locally. For most people, the API or OpenRouter is the realistic route. (Mistral's post doesn't publish hardware requirements; check the Hugging Face model card.)
- Coding-agent users: the token-efficiency claim is the part to test. Run your own agent tasks through OpenRouter before you commit.
Video
Video by Mistral on YouTube.
Video by Bijan Bowen on YouTube.
Sources
- Mistral AI, "Mistral Large 4 (Le Chonk)… open weight," 2026-10-29: https://mistral.ai/news/mistral-large-4-oss/
- Hacker News front page (Algolia API): https://hn.algolia.com/api/v1/search?tags=front_page
What is confirmed, and what is not
- ✅ Every benchmark number was copied by hand from Mistral's post on 2026-10-29 at ~11:58 PM ET. All of them are vendor-reported.
- ✅ The license wording and availability channels were checked against the same post.
- ⚠️ UNVERIFIED: the Hugging Face repo name, hardware/VRAM requirements and API pricing. None of these appear in the post.
- ⚠️ We don't name the three reasoning modes here because we couldn't read their labels reliably.
- The HN points are a snapshot and will change.
Up: AI