THE SCOREBOARD FOR WHAT'S NEXTAI + QUANTUM, SCORED. COMPARED. EXPLAINED.
AIQUANTUMSCORE.
← AI & Quantum News

AI / NEWS BRIEF

Mistral releases Large 4 weights for self-hosted deployment

Mistral has followed through on its preview promise by releasing Large 4 model weights, enabling local deployment under a modified MIT-style licence.

By AIQuantumScore · · Event 2026-10-29

The weights are now available

Mistral released the improved Large 4 weights after introducing the model in public preview earlier in October. The company describes a natively multimodal mixture-of-experts model with roughly one trillion total parameters and about 49 billion active parameters per token. Hugging Face lists official FP8 and NVFP4 variants and documents support in Transformers, vLLM and llama.cpp. Reuters separately confirmed that organizations can download and run the model independently.

Licence and hosted pricing

The weights use the Mistral Ecosystem Open-Weight License, which Mistral characterizes as a modified MIT licence. It permits use, modification, fine-tuning and commercialization, but requires a separate commercial licence for businesses offering the model as a service. Mistral also lists hosted API pricing of $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Claims still need independent testing

OUR ANALYSIS

Mistral reports substantial gains in coding, defensive cybersecurity and reasoning efficiency, and Hugging Face summarizes additional benchmark results. Those measurements depend on particular prompts, reasoning modes, runtimes and evaluation harnesses. Open weights make independent replication and safety research possible, but openness is not itself proof of capability or safety. AIQuantumScore will not assign a REVAL score without running the exact checkpoint through the benchmark.

More to follow