Paris-based Mistral AI has released the full open weights of Mistral Large 4, nicknamed “Le Chonk”, turning its flagship artificial intelligence model from an API-only preview into a system developers can download and run on their own servers. The release landed on October 29, 2026, three weeks after the model first appeared in a public preview on October 6, and it comes at a moment when the economics of large AI models are under intense scrutiny.
The release turns a preview into something developers can actually run themselves. The weights are published on Hugging Face, and the model is also served through Mistral’s own API and partner platforms, including Microsoft Foundry and Nvidia-hosted endpoints. That combination matters: the same model can now be tested as a hosted service or deployed inside an organisation’s own infrastructure, rather than only through Mistral’s preview.
Key Takeaways
- Mistral released the open weights of Large 4 on October 29, three weeks after its October 6 preview, with about 1 trillion total parameters and a 1-million-token context window.
- The model is pitched at coding and cyber defence, with Mistral reporting stronger coding scores after three more weeks of reinforcement learning.
- It is free to use, modify and build commercial products on under a modified MIT licence, except that businesses offering it as a hosted model service need a separate commercial licence from Mistral.
What Was Released
Mistral Large 4 is a mixture-of-experts model, a design in which many specialised components, or experts, are available but only a small subset is used for each token. According to the model documentation, it has roughly 1 trillion total parameters, with 128 routed experts, four active per token plus one shared expert, and about 52 billion active parameters per token including embeddings and output layers, or 49 billion excluding them. The difference is a matter of what is counted, not a different model configuration.
The model is natively multimodal, accepting text and image input, and it supports a context window of 1 million tokens, allowing it to work with very long documents and codebases in a single session. A vision encoder of about 1.6 billion parameters handles image input at variable sizes and aspect ratios. Official checkpoints are provided in FP8 and NVFP4 formats, alongside day-one quantised versions from the open-source community for running the model with lower hardware requirements.
The release turns a preview into something developers can actually run themselves.
Coding and Cyber Defence Claims
Mistral is pitching Large 4 most strongly at two jobs: writing and fixing code, and helping cyber defenders. On coding, the company reports 48 percent on the Artificial Analysis Coding Agent Index, effectively tied with China’s GLM-5.3 as the top open-weight model there, while using 47 percent fewer tokens per task. It also reports 41 percent on Terminal-Bench 4.0, a benchmark of terminal operations, up from 28 percent at preview after three additional weeks of reinforcement learning, and 19.7 percent on Cognition’s FrontierCode 2.0 benchmark.
On cyber defence, Mistral says the model scores 51 percent on the Artificial Analysis Cyber Index and that, in its own tests, it refused 96.6 percent of malicious cyber requests. It also reports 82 percent on a vulnerability reproduce-and-patch test and 93 percent on Cybench, a set of 40 challenges drawn from security competitions. These are the company’s and its partners’ numbers, not independent tests, and independent validation will take time. Mistral has also said a more aggressive penetration-testing version stays restricted to vetted partners, rather than being released openly.
- About 1 trillion total parameters, with roughly 52 billion active per token, and a 1-million-token context window.
- Terminal-Bench 4.0 score reported at 41 percent, up from 28 percent at the October 6 preview.
- Coding Agent Index score reported at 48 percent, using 47 percent fewer tokens per task than GLM-5.3.
- API pricing of $0.70 per million input tokens and $2.10 per million output tokens, roughly half the preview price.
- Modified MIT licence, with a separate commercial licence required for model-as-a-service businesses.
Licence and Pricing
The weights are published under the Mistral Ecosystem Open-Weight Licence, a modified version of the MIT licence. It permits use, modification, fine-tuning and commercialisation, with one carve-out: companies that offer Large 4 as a hosted model service need a separate commercial licence from Mistral. That distinction will matter for cloud providers and AI platforms, but not for most developers building products on top of the model.
Through Mistral’s own API, the model costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. On the Artificial Analysis Intelligence Index v5, which aggregates knowledge work, coding, science and general ability, Large 4 scores 26, the highest among Western open-weight models but fifth among open weights overall, with Chinese models such as DeepSeek V4.1 Flash and GLM-5.3 still ahead on the aggregate.
Why This Release Matters Now
The timing is significant because investors are asking harder questions about whether massive AI spending will pay off. Alphabet, Google’s parent company, reported third-quarter revenue up 25 percent to $127.58 billion, yet its shares fell sharply after the results, as covered in our report on the tech sell-off driven by AI spending fears. At the same time, cloud demand from AI infrastructure helped drive Google Cloud revenue to more than double, showing how the same earnings season can tell two different stories about AI returns.
An open-weights release speaks directly to that debate. Instead of only renting access to a frontier model, organisations can inspect, fine-tune and host it themselves, which can change the cost calculation for coding assistants, internal tools and security workflows. It also keeps the open-weights race competitive at a time when Anthropic’s reported IPO filing underlines how much value investors now attach to leading AI labs, whether their models are open or closed.
What to Watch Next
What to watch next is whether independent tests confirm Mistral’s coding and cyber-defence scores, how widely developers actually deploy the model on their own servers given its hardware demands, and how the licence carve-out for hosted model services is applied in practice. The other question is whether Large 4 narrows the gap with Chinese open-weight models on aggregate benchmarks, or whether its advantage stays concentrated in coding efficiency and cyber defence. Those answers will decide whether this release is remembered as a genuinely useful open tool or simply as another large model in a crowded field.