Mistral releases open weights for Large 4 trillion-parameter model

On October 29, 2026, French AI company Mistral released downloadable open weights for its 1.05 trillion-parameter Mistral Large 4 model, nicknamed 'Le Chonk.' Trained on 3,800 Nvidia Grace Blackwell GPUs, the model achieves 26 points on Artificial Analysis's Intelligence Index v5 at $2.06 per task—a 63% cost reduction from its preview. CEO Arthur Mensch and VP Pierre Stock emphasized that the open release offers Western enterprises sovereign cybersecurity and coding capabilities without cloud lock-in.

Model launch and architectural parameters

• ▪Mistral Large 4 utilizes a sparsely activated architecture containing 1.05 trillion total parameters and 52 billion active parameters, offering a 1 million token context capacity
• ▪Mistral AI released downloadable open weights for its 1.05 trillion-parameter AI model, Mistral Large 4
• ▪Mistral Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in European data centers following three weeks of reinforcement learning after Mistral Large 4's October 2026 preview

Intelligence Index performance and efficiency

• ▪Artificial Analysis gave Mistral Large 4 a score of 26 on Artificial Analysis's Intelligence Index v5 benchmark at high reasoning, making it the highest-scoring open-weight model developed outside China
• ▪Artificial Analysis reported that Mistral Large 4 costs $2.06 per Intelligence Index task, down 63% from $5.54 for Mistral Large 4's preview version, while consuming an average of 82,000 output tokens per task

Coding benchmarks and agent performance

• ▪In a Surge Code blind preference study reported by Mistral, Mistral Large 4 won 64% of comparisons against GLM-5.3 and Kimi K3, and 56% against DeepSeek V4.1 Flash
• ▪Artificial Analysis ranked Mistral Large 4 paired with Mistral Vibe software second among open-weight coding agents with a score of 48 on Artificial Analysis's Coding Agent Index, using 8.5 million tokens per task

Cybersecurity benchmarks and enterprise preview

• ▪Enterprise design partners, including Orange Cyberdefense, are receiving preview access to a specialized security variant of Mistral Large 4 subject to vetting and contractual safeguards
• ▪Mistral reported that Mistral Large 4 achieved a 96.6% average refusal rate on malicious cyber prompts across JailbreakBench, StrongREJECT, and AgentHarm benchmark tests
• ▪Mistral Large 4 scored 51 on Artificial Analysis's Cyber Index at high reasoning, including 79% on CyberGym-E2E-AA, 15% on DeepsecBench-AA, and 58% on CWE-Bench-AA

Sovereignty and comparison with Chinese models

• ▪Artificial Analysis reported that seven Chinese open-weight models score higher globally than Mistral Large 4 on Intelligence Index v5, led by DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32 points
• ▪Mistral Vice President of Science Pierre Stock stated that Chinese AI models offer no guarantee against malicious behavior and could serve as threat vectors
• ▪Mistral Vice President of Science Pierre Stock stated that Mistral Large 4 provides Western organizations sovereign AI capabilities, offering performance competitive with Chinese models like GLM-5.3 without risk of remote access termination

Open-weight licensing terms and conditions

• ▪The Mistral Ecosystem Open-Weight License for Mistral Large 4 contains no revenue threshold and explicitly exempts internal enterprise use and end-user products with embedded model features from separate licensing requirements
• ▪Mistral released Mistral Large 4 under the Mistral Ecosystem Open-Weight License, a modified MIT license requiring a separate commercial agreement for entities operating model-as-a-service businesses