Mistral Large 4: Open-weight AI model with one trillion parameters released

French company Mistral has officially opened the weights of the enhanced version of its flagship multimodal model, Mistral Large 4, unofficially nicknamed "Le Chonk". This massive AI system features one trillion parameters and leads open-weight solutions across major coding and cybersecurity benchmarks.

• The model features one trillion parameters and leads in FrontierCode 2.0 and Terminal-Bench 4.0 benchmarks.
• Three distinct reasoning modes are available: Low, Medium, and High for varying task complexity.
• Token consumption is reduced by 47 percent while delivering comparable performance results.
• Refusal rate on malicious prompts reaches 96.6 percent according to JailbreakBench and AgentHarm tests.

Training improvements and reasoning modes

The substantial performance boost was achieved through three additional weeks of reinforcement learning conducted on top of the initial preview version. According to Mistral, this significantly enhanced coding capabilities, defensive cybersecurity, and overall logical reasoning efficiency.

The new release introduces three configurable modes: Low for simple tasks, Medium for everyday queries, and High for complex scenarios requiring precision and multi-step analytical processing. In Surge Code blind tests, the model trailed only Claude Opus 5.5, outperforming systems like GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash.

Commercial efficiency and hosting security

A key advantage for commercial deployment is superior token efficiency. On the Artificial Analysis index, Large 4 achieves results comparable to top open models while spending 47 percent fewer tokens per task, promising significant computational savings in multi-step agentic workflows.

Despite built-in safeguards and a 96.6 percent refusal rate on malicious queries, Mistral warns that self-hosting open weights means API-level safety barriers do not transfer automatically. Teams deploying the model locally must implement their own security measures.

Geopolitical context and market positioning

The release unfolds against a backdrop of geopolitical tensions and restrictions surrounding access to American AI models. Recent temporary limitations on OpenAI and Anthropic distributions have driven higher demand for sovereign European artificial intelligence.

Mistral, which raised $3.3 billion at a $24 billion valuation, positions Large 4 as the flagship of open AI, placing particular emphasis on code and cybersecurity—areas that, according to co-founder Guillaume Lample, other laboratories do not focus on sufficiently.