Mistral open-sources Large 4, a one-trillion-parameter model trained entirely in Europe
Three weeks after its API preview, Mistral has published the weights of Mistral Large 4 under a modified MIT licence. The natively multimodal mixture-of-experts model posts the best open-weight scores on Terminal-Bench 4.0 and FrontierCode 2.0, and ships with vLLM, SGLang and NVIDIA NIM support for self-hosting.
Mistral AI has released the weights of Mistral Large 4, the one-trillion-parameter model it nicknamed Le Chonk when it entered public preview earlier in October. The open-weight version benefits from three additional weeks of reinforcement learning, and the company says the extra training lifted its Terminal-Bench 4.0 score from 28 per cent to 41 per cent while cutting output tokens by around 30 per cent.
The model is natively multimodal and uses a mixture-of-experts architecture with roughly 52 billion active parameters per token. It offers three reasoning modes, with High set as the default. Mistral reports 19.7 per cent on Cognition's FrontierCode 2.0, 84.5 per cent on BrowseComp and 97.1 per cent attack resistance on the B3 agent security benchmark, numbers that place it at the top of the open-weight field, though independent replication is still pending.
Licensing is the notable choice. The weights ship under the Mistral Ecosystem Open-Weight licence, a modified MIT licence that permits use, fine-tuning and commercial deployment, but requires a separate commercial agreement from companies that want to resell the model as a hosted service. Weights are on Hugging Face, with deployment paths through vLLM, SGLang and an NVIDIA NIM container.
Hosted pricing is aggressive at $0.70 per million input tokens and $2.10 per million output tokens, available through Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints. Mistral emphasises that the model was trained from scratch on European infrastructure, a point aimed squarely at governments and regulated industries looking for alternatives to US and Chinese providers.
Why it matters
A frontier-class open-weight model with permissive terms changes the economics for anyone who must run AI inside their own data centre. It also raises the bar for other open releases, which now have to compete on agentic coding and security benchmarks rather than raw chat quality.
Indicops view
Indian ministries and public-sector banks routinely ask for models they can host on sovereign or on-premise infrastructure. Large 4 is a serious candidate for the heavy end of that spectrum, particularly for coding assistants and document-heavy agents, although its size means it needs real GPU capacity. We advise pairing it with a compact Indic model for language coverage and testing the licence terms carefully where a system integrator will operate the service on a client's behalf.
More in Open Source
Back to the newsroomHugging Face's chief vows to fight on for open source as Nvidia's $12.9 billion bid faces uncertainty
After Jensen Huang's exit as Nvidia CEO, Hugging Face chief Clement Delangue said the company will keep fighting for open-source AI harder than ever. Nvidia's planned $12.9 billion purchase of the platform is scheduled to close in the first half of 2027 and has drawn US Justice Department questions.
Nvidia donates OpenShell agent runtime to CNCF and previews GPU Kubernetes tooling
Nvidia has donated OpenShell, an open-source runtime that applies policy controls to autonomous AI agents, to the Cloud Native Computing Foundation. At KubeCon North America it also put new GPU Kubernetes tools and a fast inference-restore feature into public preview.
Baidu releases ERNIE 5.5 Flash Base, a 280-billion-parameter open-weight model with a restrictive licence
Baidu has published ERNIE 5.5 Flash Base, a 280B-parameter mixture-of-experts checkpoint with 14B active per token, trained on 33.5 trillion tokens. Weights are downloadable, but the licence excludes the EU, UK and South Korea and limits commercial use above $50 million revenue.