Mistral Opens the Weights of Its Trillion-Parameter Large 4: What ‘Le Chonk’ Changes for Open AI
Key takeaway: Mistral AI released the open weights of Mistral Large 4, nicknamed “Le Chonk,” on October 29, 2026, three weeks after its October 6 public preview. The trillion-parameter model is built for agentic coding and cybersecurity, and anyone can download and run it — except businesses that want to sell it as a hosted service, which need a separate commercial licence.
Mistral AI, the Paris-based lab, has opened the weights of its largest model yet. The release, announced on October 29, 2026, puts a 1-trillion-parameter, natively multimodal model into the hands of developers, researchers, and companies that want to run a frontier-class system on their own infrastructure rather than through someone else’s API. The weights are available on Hugging Face, and the model is also served through Mistral’s own API, Microsoft Foundry, NVIDIA-hosted endpoints, and OpenRouter.
This matters because open-weight releases at this scale are rare, and because Mistral is aiming squarely at two of the most commercially valuable jobs in AI right now: writing and reviewing code, and defending computer systems. Understand it, then use it: if you build software or run infrastructure, Large 4 is a new option to test — but its licence has a catch worth reading before you build a business on it.
What Mistral actually released
Mistral Large 4 is a mixture-of-experts (MoE) model. Instead of activating its full parameter count for every word it generates, it routes each token to a small subset of specialist sub-networks, or “experts.” That design is what makes a trillion-parameter model practical to run at all.
| Specification | Mistral Large 4 |
|---|---|
| Total parameters | About 1 trillion (1.05 trillion per some reports) |
| Active parameters per token | About 52 billion, or 49 billion excluding embeddings and output layers |
| Architecture | Mixture-of-experts: 128 routed experts, 4 active per token, plus 1 shared expert |
| Context window | 1 million tokens |
| Inputs | Text and images (natively multimodal) |
| Languages | Native fluency in more than 160 languages, per Mistral |
| Checkpoints | FP8 and NVFP4, with day-one GGUF quantizations from the community |
| Reasoning modes | Three modes — low, medium, and high — with high as the default |
One detail in the reporting deserves a note: sources differ slightly on the active parameter count — 49 billion versus 52 billion. Both figures describe the same model; the difference is whether embeddings and output layers are counted. The model card figure most often cited is 52 billion active parameters per token.
The released checkpoint is not identical to the October 6 preview. Mistral says it reflects three more weeks of reinforcement learning, and the benchmark movement below shows why the company bothered: its Terminal-Bench score jumped substantially in that window.
How Large 4 compares on coding benchmarks
Benchmark scores are vendor-reported unless noted, and they should be read that way — but the pattern across independent coverage is consistent: Large 4 leads open-weight models on coding tasks while still trailing the best closed models by a wide margin.
| Benchmark | Large 4 | Closest rival |
|---|---|---|
| Terminal-Bench 4.0 | 41% (up from 28% at preview) | GLM-5.3 at 40% |
| Artificial Analysis Coding Agent Index | 48%, using 47% fewer tokens per task than GLM-5.3 | GLM-5.3, effectively tied |
| Cognition FrontierCode 2.0 | 19.7% (vendor-reported) | DeepSeek V4.1 Flash at 19.3%; GLM-5.3 at 17.8% |
| Artificial Analysis Intelligence Index v5 | 26 points — top Western open-weight model, fifth among open weights overall | Chinese open models, including DeepSeek V4.1 Flash and GLM-5.3, ahead on aggregate |
| Claude Opus 5.5 (closed-model reference) | — | 62 points on the Coding Agent Index; 46% on FrontierCode 2.0 |
The honest reading: Mistral can fairly claim the best open-weight coding model on these measures, but the gap to the closed frontier remains large. Claude Opus 5.5 scores 62 on the same Coding Agent Index where Large 4 scores 48, and 46% on FrontierCode 2.0 against Large 4’s claimed 19.7%.
There is also a caution flag on that FrontierCode figure. One independent review that cross-checked Mistral’s numbers against benchmark owners’ boards found Artificial Analysis’s own run of FrontierCode 2.0 at 14.96%, not 19.7%, and noted that Mistral did not name the harness behind its figure. Treat the 19.7% as a vendor claim until an independent run confirms it.
What it costs to use Large 4
Through Mistral’s API, Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. That is roughly half the preview’s list price of $1.36 and $4.18 per million input and output tokens — a cut that makes the model materially cheaper to test than it was three weeks ago.
Self-hosting is a different calculation. A trillion-parameter model, even a sparse MoE, is not a laptop workload. Early community deployment notes put the standard version at eight NVIDIA B300 GPUs, with a compressed version needing four GB300s. For most teams, the API or a partner-hosted endpoint (Microsoft Foundry, NVIDIA NIM) is the realistic route; self-hosting makes sense for organizations with strict data-sovereignty requirements or existing GPU infrastructure.
The licence: open weights, with one big carve-out
Large 4 is open-weight, not open source. It ships under the Mistral Ecosystem Open-Weight License, a modified MIT licence that permits free use, modification, fine-tuning, and commercialization — with one exception: any business that offers Large 4 as a hosted model service (Model-as-a-Service) needs a separate commercial licence from Mistral first.
In plain terms: you can download the weights, fine-tune them, and sell products built on them. You cannot resell the model itself as an API without Mistral’s permission. Independent coverage has also noted that the announcement does not clearly define where a SaaS product wrapping the model ends and a Model-as-a-Service business begins — a grey area any startup building on Large 4 should get legal advice on before launch.
The cyber angle carries its own restriction. Mistral pitches Large 4 as its leading model for cyber defence, citing 51% on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests in its own tests. But a more aggressive penetration-testing variant stays restricted to vetted partners only — the public weights are the defender-oriented version, not the unrestricted offensive one.
What you need before you try it
For API use, the prerequisites are minimal: a Mistral API account or access through OpenRouter, Microsoft Foundry, or an NVIDIA endpoint. For self-hosting, you need GPU infrastructure at the scale above, plus deployment tooling such as vLLM or SGLang, both of which the weights support.
The main risk is not technical but legal and evaluative. The licence carve-out can bite a hosting business, and the benchmark figures — particularly FrontierCode — are partly vendor-reported and partly disputed. Test the model on your own tasks before committing: coding benchmarks measure general ability, not whether it handles your codebase, your stack, or your security policies.
Knowns, unknowns, and what to watch
Known: the weights are public, the API price is roughly half the preview price, and the model leads open-weight rivals on coding benchmarks while trailing closed leaders.
Unknown: whether independent evaluations will confirm the FrontierCode claim, where exactly the licence draws the line between a SaaS product and a Model-as-a-Service business, and whether the EU AI Act’s open-source exemption applies to this licence — a question Mistral’s own documents do not settle.
What to watch: independent benchmark runs in the coming days, any clarification from Mistral on the licence boundary, and rival releases — Moonshot is reported to be planning open weights for Kimi K3.1 Pro by November 4, 2026, which would immediately test Large 4’s claim to the open-weight coding crown.
For readers tracking the open-model race, the bigger story is structural: a European lab trained a trillion-parameter model from scratch on its own European data centers and gave the weights away, betting that sovereignty, inspectability, and a low API price matter more than keeping the model behind a wall. Whether developers agree will show up in downloads, fine-tunes, and real deployments — not in launch-day benchmarks.
Sources
- The Decoder: Paris-based Mistral releases “Le Chonk,” a strong coding model that still trails China overall
- Unite.AI: Mistral’s 1-trillion-parameter Le Chonk checkpoint lands on Hugging Face
- TestingCatalog: Mistral opens Large 4 weights for coding and security
- Skimwire: Mistral Large 4 open weights — specs, license, cyber claims
- The Neuron: Everything that happened in AI today, Friday, October 30, 2026
Related on USABite: Alphabet and Microsoft both rode the AI boom — only one got rewarded and AI governance business documents explained.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4 is Mistral AI’s flagship model, nicknamed “Le Chonk.” It is a 1-trillion-parameter, natively multimodal mixture-of-experts model built for agentic coding and cybersecurity, with a 1-million-token context window and fluency in more than 160 languages.
When did Mistral release the Large 4 open weights?
Mistral released the open weights on October 29, 2026, three weeks after the model’s public preview on October 6, 2026. The released checkpoint reflects three additional weeks of reinforcement learning, which lifted its Terminal-Bench 4.0 score from 28% to 41%.
Is Mistral Large 4 really open source?
No. Large 4 is open-weight, not open source. It uses the Mistral Ecosystem Open-Weight License, a modified MIT licence that allows free use, modification, fine-tuning, and commercialization, but requires a separate commercial licence for any business offering the model as a hosted service.
How much does Mistral Large 4 cost through the API?
Through Mistral’s API, Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. That is roughly half the preview’s list price of $1.36 and $4.18 per million input and output tokens.
How does Mistral Large 4 compare with closed models like Claude?
Large 4 leads open-weight models on coding benchmarks, scoring 48% on the Artificial Analysis Coding Agent Index and 41% on Terminal-Bench 4.0. But it still trails the closed frontier: Claude Opus 5.5 scores 62 on the same Coding Agent Index and 46% on FrontierCode 2.0, against Large 4’s vendor-claimed 19.7%.
