The Weights Are Out: A Trillion-Parameter Model Anyone Can Download
French AI company Mistral has announced that the full weights of its flagship model, Mistral Large 4 (nicknamed Le Chonk), are now publicly available for download. The one-trillion-parameter, natively multimodal model debuted in public preview earlier this month; after just three weeks of additional reinforcement learning, it is being released openly in substantially stronger form — letting any company, government agency, or developer download it and run it on their own infrastructure.
Unlike closed models, open-weight models hand ownership back to the user: organizations keep full control of their data flow and can customize and fine-tune freely, without depending on any cloud API. Mistral stresses that Mistral Large 4 was trained entirely on European infrastructure with a clear lineage from inception to release — a trusted foundation for what it calls Sovereign AI.
Coding Crown: Best Open-Weight Model
The headline story of this release is coding performance. On Cognition’s private FrontierCode 2.0 benchmark — which assesses whether an AI-generated change would actually be merged by a repository maintainer — Mistral Large 4 leads with 19.7%, ahead of GLM-5.3, Kimi K3, DeepSeek V4.1 Flash and other open-weight rivals, taking the top spot among open models.
On Terminal-Bench 4.0, the model jumped 13 percentage points over its preview to 41%, while cutting output tokens by 30% — performance and efficiency improving simultaneously. On the Artificial Analysis Coding Agent Index, it scores 48%, effectively tied with the leading open-weight model GLM-5.3 while using 47% fewer tokens per task on average, dramatically lowering the cost of running coding agents at scale.
The model now ships with three reasoning modes (Low / Medium / High), letting developers balance performance against cost for each workload.
Cybersecurity Is the Hidden Ace
Beyond coding, Mistral Large 4’s strength in defensive cybersecurity deserves attention. It scores 51% on the Artificial Analysis Cyber Index, which measures a model’s ability to discover vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality.
Mistral has been working with leading cybersecurity vendors to validate the model’s vulnerability-reproduction and patching capabilities in real-world environments. Crucially, it handles these legitimate defensive tasks without unnecessary refusals — a practical pain point for enterprise security teams. The model supports a 1 million-token context, with API pricing at $0.70 per million input tokens.
Europe’s Sovereign AI Bet — and Its Caveats
Mistral positions itself as the definer of the “Sovereign AI” category: in an era when two superpower blocs dominate frontier models, Europe needs an AI option it can fully control. Trained in Europe, open-weight, and self-hostable, Mistral Large 4 is the centerpiece of that strategy.
Open weights, however, do not mean zero barriers. Self-hosting this trillion-parameter model requires roughly 1.06 TB of weight files and serious compute; the license also requires a separate commercial agreement for some hosted AI services. Third-party evaluator Artificial Analysis gives it 26 points — eighth among open models, ahead of Meta but still behind seven models from China.
Conclusion: Europe’s Answer in the Open-Source Arena
The open-sourcing of Mistral Large 4 marks Europe’s most confident move yet in the frontier AI race. It tops the open-weight charts in the two categories enterprises need most — coding and defensive cybersecurity — while its transparent lineage and self-hosting capability make it highly attractive to data-sovereignty-conscious governments and enterprises. For developers, there has never been a better time to adopt a self-hosted frontier model; for the industry, the gap between open and closed camps is closing fast — and this race is only entering its second half.
