Mistral Large 3: 128 MoE experts, 512k context, open weights

Mistral AI announced Mistral Large 3, an open Mixture-of-Experts model with 128 experts (8 active per token), 420B total and 68B active parameters. It offers a 512,000-token context window and scores 53.8% on SWE-bench Verified and 65.2% on Terminal-Bench 4.0. Weights ship under a permissive license and via La Plateforme.

• MoE design: 128 experts, 8 activated per token, 420B total parameters
• 68B active parameters at decoding with a 512,000-token context window
• 53.8% on SWE-bench Verified and 65.2% on Terminal-Bench 4.0
• Open FP8/BF16 weights for on-premise and air-gapped deployment