Mistral AI has made the weights of its flagship model, Mistral Large 4 (internal name Le Chonk), publicly available. This is the largest open multimodal MoE model trained from scratch entirely in Europe: 1 trillion parameters, with 52B active per token and a 1M context window. After the public preview, the model was fine-tuned using RL for coding and cybersecurity. The weights are published on Hugging Face under a modified MIT license that permits self-hosting but requires a separate license for commercial use of the model as a service. Mistral Large 4 can be tried without local installation via the Mistral API, OpenRouter, and other ready-made channels.

image
image
image

What happened

Mistral placed the weights for Mistral Large 4 in the mistralai/Mistral-Large-4-1T-A52B repository on Hugging Face and published a technical announcement. The model is built as a mixture-of-experts: 128 routing experts plus one shared, with 52B active parameters per token out of a total of one trillion. Training from scratch took place on 3,800 NVIDIA Grace Blackwell accelerators in European data centers, and after the preview, the model was fine-tuned via RL in the areas of coding and cybersecurity. According to the announcement, it scored 48% on the Artificial Analysis Coding Agent Index — on par with GLM-5.3 but with 47% lower token consumption per task — and 51% on the Artificial Analysis Cyber Index, which covers vulnerability search, reproduction, and patching. Additional security figures: 96.6% refusal rate on malicious requests and 97.1% on B3 Agent Security. To manage generation costs, three reasoning modes are provided — Low, Medium, and High (enabled by default). The license is named the Mistral Ecosystem Open-Weight License and is a modified version of MIT, where commercial use of the model as a service (Model-as-a-Service) requires a separate license from Mistral.

Context

The release is significant not in itself, but as a fork in the open-weight landscape: Mistral has for the first time released a flagship-class multimodal model into the open, trained entirely on European infrastructure, and accompanied it with launch recipes. The architectural logic here is sparse scaling: thanks to routing across 128 experts, only 52B of the trillion parameters work per token, allowing for moderate generation costs despite the large total volume, while the 1M context expands scenarios for agents and working with large documents. The three reasoning modes reflect the understanding that the volume of computation at the generation stage determines not only the quality of the answer but also its cost. Mistral's bet on this release is Sovereign AI: companies and government agencies should be able to inspect, fine-tune, and deploy the model on their own infrastructure without dependence on someone else's API. Notably, the chosen direction of specialization is defensive scenarios, meaning the model was deliberately improved on finding and closing vulnerabilities with a high refusal rate on malicious requests. For a correct assessment, it is worth remembering the limitations of the comparison set: the key comparison on token efficiency was made only with one model, GLM-5.3.

Why this matters for the industry

For companies and developers, open weights provide a predictable alternative to APIs: the model can be inspected, fine-tuned, and deployed on their own infrastructure, and the barrier to entry is lowered by ready-made deployment recipes. The economics are transparent from day one: via the Mistral API, the price is $0.70 per million input tokens, $0.07 with input caching, and $2.10 per million output tokens. The appearance of a Europe-trained frontier flagship at this price lowers the lower bound of the cost of strong inference and pushes toward the commoditization of access to top models. For startups, the signal is twofold: thin wrappers over third-party APIs lose rent, but space opens up for products that build value on self-hosting, customization, and domain scenarios. A separate growth point is the defensive security profile: high results on the Cyber Index combined with a low false positive rate on legitimate tasks make the model a candidate for internal code auditors and primary triage of vulnerability reports. However, cloud providers need to keep the licensing restriction in mind: commercial deployment of the model as a service requires a separate license from Mistral, so reselling the weights as their own API for free is not possible.

Why this matters for users

The model can be tried right now without local installation: via the Mistral API under the identifier mistral-large-4, via OpenRouter under mistralai/mistral-large-4-0, in the Vibe Code environment, or through optimized NVIDIA NIM containers. For running on their own hardware, Mistral has released ready-made vLLM (vllm/vllm-openai:mistral-large-4) and SGLang images, with the combination with DSPark speculative decoding claiming up to 1.9x throughput. Switching the reasoning mode allows balancing answer quality and generation cost, which is especially noticeable in agents with many steps. The 1M context and support for 160+ languages make the model a convenient base for assistants over large documents, code review automation, and processing requests in different languages. It is logical to build initial prototypes around code auditing and vulnerability report analysis — these are the directions the model was fine-tuned for after the preview. A practical bonus is that the same client code works almost unchanged both via API and with a self-hosted deployment, so the model can be safely tested on your own tasks before any infrastructure commitments.

What is still unknown / limitations

The main caveats concern performance figures: 48% on the Coding Agent Index and 51% on the Cyber Index rely on Artificial Analysis estimates from the vendor's announcement — these are currently claims, not independent measurements, and there are no public test sets, evaluation protocols, or third-party replications. The claimed token savings have only been verified in comparison with one model, GLM-5.3, so extrapolating them to other model pairs is premature. For self-hosting, there are currently no public data on actual hardware requirements, latencies, and real operating costs beyond the ready-made images. Expectations that quantizations, forks, and fine-tuned variants will quickly form around the open weights remain predictions, not accomplished events; the announced efficiency should be verified on your own workload before incorporating figures into product plans.

Sources

Author

Look at AI, editorial team