Mistral Large 4 Achieves Coding Performance Similar to GLM-5.3 with 1T MoE Model

• Mistral AI released the weights for Mistral Large 4, a 1T parameter MoE architecture, on Hugging Face on October 29
• Recorded performance similar to GLM-5.3 (49 points) with a Coding Agent Index score of 48, while using approximately 47% fewer tokens on average per task
• Supports 1M token context and image input, with an FP8 checkpoint size of approximately 1.06TB
• Use of a dedicated vLLM Docker image is mandatory, and verified hardware configurations such as 8 B300s and 16 H200s are provided

Benchmarks and Performance

• Ranked 3rd with a Cyber Index score of 51%, following DeepSeek-V4.1-Flash (57%) and MiMo-V2.6-Pro (56%)
• Claimed to be the highest score among Western open-weight models and 5th among open-weight models overall with an Intelligence Index v5 score of 26
• Claimed up to 1.9x throughput improvement when applying DSpark for speculative decoding, with measured generation speeds of 120–138 tokens per second on conversational prompts

Serving and License

• Applies the Mistral Ecosystem Open-Weight License, allowing free internal use and product embedding
• A separate contract with Mistral is required for commercial use when providing MaaS (Model-as-a-Service)
• API pricing is $0.70 per 1 million input tokens and $2.10 per 1 million output tokens