Mistral Small 4

• 🆕 Unifies instruct, reasoning, and coding into one model
• 🧠 119B-parameter MoE, only ~6.5B active per token
• 🔧 128 experts, 4 active per token
• 📏 Native 256K-token context window
• 👁️ Accepts text and image input, text output
• 🎯 Configurable reasoning effort, toggle fast vs. thinking mode
• 🔒 Apache 2.0 license, open weights
• 💬 Native function calling and JSON output for agentic use

Mistral AI

Mistral AI is a French artificial intelligence company headquartered in Paris, founded in 2023. The company focuses on developing large language models offered under both open-weight and proprietary licenses. Mistral AI has quickly risen to prominence in the…

Mistral Small 4, released March 16, 2026, is Mistral AI's open-weight hybrid model that consolidates three previously separate model families—Instruct, Reasoning (formerly Magistral), and Devstral coding—into a single checkpoint. It uses a Mixture-of-Experts architecture with 119 billion total parameters spread across 128 experts, activating only 4 experts (about 6.5 billion parameters) per token, giving it the inference profile of a much smaller dense model. It accepts text and image input, supports a 256K-token context window, and exposes a configurable reasoning-effort parameter that toggles between fast instant replies and a slower thinking mode.

This marks a substantial change from its same-family predecessor, Mistral Small 3.2 24B Instruct, a 24B dense instruction model from January 2026. Where Small 3.2 focused on text instruction following, Small 4 moves to a sparse MoE design, adds native vision, integrates dedicated reasoning and agentic-coding behavior, and expands deployment to enterprise-scale tasks.

Mistral distributes Small 4 under the Apache 2.0 license with multiple checkpoints, including an NVFP4 4-bit quantized version and a trained Eagle head for speculative decoding. Mistral positions it for chat assistants, coding, agentic workflows, and reasoning tasks, with native function calling and JSON output.