Mistral releases open weights for Mistral Large 4 and cuts API prices Other

Mistral AI published open weights for Mistral Large 4 on Hugging Face and reduced its commercial API pricing following three weeks of post-announcement fine-tuning.

The released 1-trillion parameter Mixture-of-Experts model features enhanced performance in agentic coding, reasoning, and defensive cybersecurity workflows, accessible under the Mistral Ecosystem Open-Weight License.

What it is

• Granular Mixture-of-Experts architecture. Mistral Large 4 contains approximately 1.05 trillion total parameters, activating between 49 and 52 billion parameters per token via 128 routed experts.
• Post-announcement tuning. Mistral spent three weeks refining checkpoints after the initial preview, yielding measurable gains in vulnerability reproduction, software patching, and multi-turn coding benchmarks.
• Open weights availability. Official FP8, NVFP4, and community GGUF quantized checkpoints are now hosted on Hugging Face for self-hosted deployment.
• Price reduction on commercial API. Mistral lowered hosted API rates to $0.70 per 1M input tokens and $2.10 per 1M output tokens.
• Embedded speculative decoding. The architecture incorporates an integrated draft head to accelerate token generation throughput during long-context inference.

Pricing and specifications

Sources

• Mistral Large 4 on Hugging Face
• Mistral AI News