Mistral releases open-source weights for 1 trillion parameter Large 4 model

From preview to public weights

Mistral AI has officially released the weights for its flagship model, Large 4, also known as Le Chonk. This move transitions the model from a public preview accessible only via API to a state where the underlying files are available on Hugging Face. The release allows developers to download, modify, and retrain the model for their own specific needs.

The model is significantly large, boasting 1.05 trillion parameters. Despite this scale, it is designed to be efficient by using a Mixture of Experts architecture. This technical approach means the model only activates a portion of its parameters—roughly 49 billion to 52 billion—to process each individual token, which helps save on computing power.

Mistral spent three weeks conducting additional reinforcement learning after the initial preview launch. This training phase was specifically designed to sharpen the model's capabilities in programming, cybersecurity, and agent-based tasks. The company plans to release a full technical report regarding these results at a later date.

Coding and security benchmarks

The performance of Large 4 has been validated through several independent and internal tests. In the Terminal-Bench 4.0, which measures how AI agents handle programming via a terminal, the model's score rose by 13 percentage points to 41% compared to the preview version. It also achieved a 30% reduction in output token usage during these tasks.

In the Artificial Analysis Coding Agent Index, Large 4 scored 48%, nearly matching the performance of China's GLM-5.3. During these tests, Mistral reported that the model used 47% fewer tokens on average for similar tasks. In a blind comparison by Surge Code, the model was preferred over several competitors, including DeepSeek V4.1 Flash and Kimi K3, and was only outperformed by Anthropic’s Claude Opus 5.5.

Mistral also prioritized cybersecurity, with the model scoring 51% on the Artificial Analysis Cyber Index. This index tracks how well an AI can identify and fix vulnerabilities in code. Furthermore, the model successfully rejected malicious requests related to cyberattacks in 96.6% of cases. In tests for instruction injection, its resilience improved from 93.3% in the preview to 97.1%.

Licensing and accessibility

The model is distributed under the Mistral Ecosystem Open-Weight license, which is a modified version of the MIT license. This means that companies can use, modify, and integrate Large 4 into commercial products for free. However, there is a specific restriction for companies that want to sell access to the model as a cloud service, which requires a separate negotiation with Mistral.

While the model weights are public, the specific data used to train the neural network remains private. Users who do not wish to host the model themselves can still access it through Mistral Studio, the Mistral API, or OpenRouter. For those choosing the API route, Mistral has set prices of $0.7 per million input tokens and $2.10 for output tokens.

The model was trained using 3,800 Nvidia Grace Blackwell accelerators located in Mistral's European data centers. This infrastructure allowed for the creation of a model that supports over 160 languages and can handle a context window of up to 1 million tokens. It also supports multimodal inputs, allowing it to process both text and images.

Corporate and safety guardrails

Mistral has taken a nuanced approach to the security risks associated with such a powerful model. They acknowledged that high-performing AI could potentially be used to facilitate real-world cyberattacks. To manage this, they have created a special version of Large 4 specifically for advanced security tasks like penetration testing.

This specialized version is not available to the general public. It is only accessible to corporate partners who undergo a thorough vetting process and sign a formal contract. Major companies including Booz Allen Hamilton and CrowdStrike have already participated in testing this version in practical environments.

Several other organizations have played a role in the model's development and validation. Cisco and Cloudflare supported the release, while the European Union Agency for Cybersecurity (ENISA) conducted tests on the model. Their testing focused on the model's ability to perform log analysis, code auditing, and vulnerability detection.