Mistral Releases Open Weights for Its 1 Trillion-Parameter Large 4 Model

The model has 1 trillion parameters and is natively multimodal: it takes text and images as input and returns text and code. According to the Hugging Face repository name, Mistral-Large-4-1T-A52B, it activates 52 billion parameters for each task. The Next Web, which reported the release on October 29, said a quantised version runs on four Nvidia B200 or B300 GPUs. Mistral has nicknamed the model "Le Chonk".

Mistral released Large 4 as a public preview earlier in October. The Next Web reports that the company then spent about three more weeks on reinforcement learning, guided by customer feedback, before publishing the weights.

Licence and Pricing

The weights come under the Mistral Ecosystem Open-Weight licence, which Mistral calls a modified MIT licence. The company says the model can be used, modified, fine-tuned and commercialised for free. The one exception is businesses that run a model-as-a-service offering, which need a separate commercial licence from Mistral.

For companies that would rather not host it themselves, the model is available through Mistral Studio, the Mistral API and OpenRouter. It is also offered on Nvidia-hosted endpoints and in an Nvidia NIM container, and on Microsoft Foundry with private deployment options on Azure. Mistral's announcement lists API pricing of $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The Next Web's report said pricing had not yet been published.

Performance Claims

Mistral is selling Large 4 mainly on coding, agentic work and cyber security. All of the benchmark scores below come from the company, and The Next Web noted that the Terminal-Bench figure has not been independently replicated.

On Terminal-Bench 4.0, an agentic coding test, Mistral says the model scores 41%, up from 28% at the preview stage. The Next Web reported scores of 40% for GLM-5.3, 31% for DeepSeek V4.1 Flash and 21% for Kimi K3. Mistral says the model scored 48% on the Artificial Analysis Coding Agent Index, roughly level with GLM-5.3, while using about 47% fewer tokens per task. It also claims 19.7% on Cognition's private FrontierCode 2.0 benchmark, which it describes as the best score among open-weight models.

In security, Mistral reports 51% on the Artificial Analysis Cyber Index. It says the model refused malicious cyber requests 96.6% of the time on average, and resisted 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% at preview. In human preference tests on code from Surge, the company says the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked behind only Anthropic's Claude Opus 5.5.

Why It Matters

Mistral is pitching the release as a way for enterprises and governments to keep control of their AI. In The Next Web's report, Pierre Stock, Mistral's vice-president of science, said "AI is going to be like electricity" and that no company should depend on a provider that can withdraw access from one day to the next. Chief executive Arthur Mensch told the outlet the company has raised about $6 billion in three years, far less than rival labs.

The Next Web reported that the model was trained on 3,800 Nvidia Grace Blackwell GPUs in Europe, with reinforcement learning on Mistral's own compute in France drawing about 10 megawatts. Mistral plans smaller distilled versions in the coming months, including one that runs on a laptop, but has not given dates.

The release gives buyers a frontier-scale model from a lab outside both the US and China that they can inspect and run inside their own data centres. That is likely to interest regulated industries and governments, including in the Gulf, that are building sovereign AI capacity.