Mistral AI has released the weights of Mistral Large 4, nicknamed "Le Chonk," on 29 October 2026. It is a 1-trillion-parameter model trained in Europe, and Mistral says it is the strongest open-weight model for coding. Anyone can now download it. Running it, however, is a different matter.
What is in the release
The weights are on Hugging Face in FP8 and NVFP4 formats, and the full repository is roughly 1.06 terabytes. According to coverage of the model card, Large 4 is a mixture-of-experts model with about 49 billion parameters active at a time and a context window of one million tokens. It accepts text and images. Mistral's own API prices it at $0.70 per million input tokens and $2.10 per million output tokens.
The licence, and the catch
Large 4 ships under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT licence. You can use, modify, fine-tune and commercialise it. The single condition is that businesses running a Model-as-a-Service offering, meaning selling access to the model to others as a hosted service, must obtain a separate commercial licence from Mistral. It is not a standard open-source licence, so teams building products should read the short licence file carefully.
Mistral's performance claims
- Terminal-Bench 4.0: 41%, up 13 points from the public preview, with 30% fewer output tokens.
- Artificial Analysis Coding Agent Index: 48%, roughly level with China's GLM-5.3 while using 47% fewer tokens per task.
- B3 Agent Security Benchmark: resists 97.1% of attacks, up from 93.3% in the preview.
- A 96.6% average refusal rate on malicious cyber requests.
These figures come from Mistral and have not all been independently confirmed. Independent index scores for the preview placed it behind leading Chinese open models such as GLM-5.3 and Kimi K3 on overall intelligence, so "best open-weight model" holds for coding and, in Mistral's framing, for Western models rather than across the board.
Sovereignty is the pitch
Mistral is positioning Large 4 for organisations that want to run AI on their own servers rather than send data to a US cloud: European-trained, self-hostable and customisable. The company also warns that the model's cyber capabilities could aid attacks, and says it red-teamed it with security partners before release.
What it means in practice
A model of this size needs serious GPU hardware. For most developers and startups, including across Africa, the realistic routes are Mistral's API or a hosting provider that serves it, not a home download. That makes the licence's hosting clause more relevant than it first looks. The release matters most for governments, banks, telecoms and large firms with strict data-residency needs, and for researchers who want a frontier-scale open model to study and fine-tune.
It also shows how crowded the open-weight race has become. The model on top of a particular chart changes within weeks, and closed models from OpenAI, Google and Anthropic still lead overall. For anyone building with AI, the sensible approach is to test Large 4 on your own tasks and compare cost and quality with what you use today.
Sources: Mistral AI announcement (29 October 2026); DEV Community; Data Studios; Resultsense; Beam.ai; Kingy AI.
