Mistral 3 at a glance

What to consider

• Mistral 3 is a December 2025 family, not a guarantee that every member is the newest or strongest model in Mistral's current catalog.
• Mistral says Large 3 is not a dedicated reasoning model and can lose to specialized reasoning systems on strict reasoning tasks.
• The Large 3 model card says it trails vision-first models on some multimodal work.
• Large 3 is difficult to self-host: even the optimized NVFP4 checkpoint is positioned for an eight-A100 or eight-H100 node.
• Smaller Ministral variants trade capacity and difficult-task reliability for lower latency and deployment cost.
• A stated 256K context window does not guarantee equal recall, precision, or instruction following across the entire window.
• Reasoning variants can use more tokens and time while still producing an incorrect result.
• Quantization and alternative runtimes can change output quality, throughput, memory use, and tool behavior.
• Open weights do not mean the full training dataset and training process are reproducible.
• Apache 2.0 does not eliminate obligations involving input data, privacy, output rights, trademarks, regulated uses, or local law.
• Function calling is probabilistic; arguments must be schema-validated and tools should have minimal permissions.
• Self-hosting transfers patching, access control, abuse prevention, monitoring, uptime, and incident response to the operator.
• API and third-party-provider data terms differ from local deployment and must be reviewed independently.
• Vendor benchmarks and leaderboard positions may not predict quality on a specific language, image type, domain, or prompt style.
• Safety-sensitive applications need separate moderation, evaluation, and human escalation; model access alone is not a complete safety system.