Mistral Large 4: "Le Chonk"
3 hours ago
- Mistral launches public preview of Mistral Large 4 (ML4), nicknamed 'le Chonk', with API access on Mistral Studio and open weights arriving by end of month.
- ML4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters, combining instruction-following, reasoning, and agentic capabilities.
- It achieves frontier-level performance in coding, agentic workflows, cybersecurity, finance, law, and multimodal understanding, outperforming other Western open-weight models.
- The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, supporting AI sovereignty and European data controls.
- Exceptional cybersecurity results include 82% on a real vulnerability reproduction test and 93% on Cybench, avoiding refusal issues seen in closed models.
- Agentic coding scores include 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, with strong human-evaluated coding quality.
- Multimodal strengths include visual grounding, satellite imagery analysis, technical drawing inspection, and document reasoning, with performance surpassing some closed models.
- Strong scientific and math capabilities include one-shot Hartree–Fock simulation generation and state-of-the-art results on SciCode-Verified among open-weight models.
- Robust safety features include high resistance to indirect prompt injections (93.3% on Lakera B3) and high refusal rates for malicious requests.
- Mistral plans rapid further improvements through scaled reinforcement learning and expanded compute, with full model details, benchmarks, and methodology to follow.
- Pricing is set at $1.36 per million input tokens and $4.18 per million output tokens.