Mistral - Frontier AI. In Your Hands.

Welcome to the official Mistral Hugging Face organization. Here you'll find our full model lineup: generalist models — state-of-the-art, open-weight workhorses for reasoning, code, multimodal understanding, and agentic workflows — alongside specialist models built with Forge, our platform for training custom AI grounded in proprietary knowledge across pre-training, post-training, and reinforcement learning. Browse our collections below to find the right model for your use case.

General Purpose & Reasoning

Frontier

• Large 4: A state-of-the-art, open-weight, general-purpose multimodal model with a granular Mixture-of-Experts architecture. 1T total / 52B active parameters, natively multimodal (text + vision), 1M context window, fluent in 160+ languages. State-of-the-art among open-weight models on cybersecurity, finance, and law.
• Medium 3.5: A balanced, multimodal frontier model optimized for cost-performance across enterprise workloads. Strong reasoning, instruction following, and tool use at a fraction of flagship compute cost.
• Small 4: An efficient model delivering frontier-adjacent performance for high-throughput, latency-sensitive deployments. Ideal for edge inference, high-volume APIs, and cost-constrained production environments.

Edge Models

• Ministral 3: A collection of edge models (3B, 8B, 14B) with Base, Instruct, and Reasoning variants. All with vision capabilities. Built for on-device and on-prem deployment where control, privacy, and low latency matter.

Specialized Models

• Voxtral: Mistral AI audio models. Real-time speech-to-text, text-to-speech, and voice agent capabilities with multilingual support, streaming inference, and native audio understanding.
• Timestral: Time series foundation models for zero-shot probabilistic forecasting. Built with Forge for domain-specific temporal reasoning across finance, energy, and operations.
• Shieldstral: A policy-adaptive multimodal safety classifier built with Forge. Evaluates content against custom policies with a single yes/no token, enabling real-time guardrails for enterprise AI deployments.
• Leanstral: The first open-source code agent for Lean 4, the proof assistant used in formal mathematics and software verification. 6B active parameters, designed for realistic formal repositories.

New checkpoints will be released in this organization. We strive to ensure their correctness to the best of our ability. For more checkpoints also see mistral-experimental.