Mistral Small 4

Mistral Small 4 is a landmark open-source AI model released in March 2026 under the Apache 2.0 license — meaning full commercial use, fine-tuning, and redistribution with zero royalties. It immediately topped open-source reasoning benchmarks on MMLU-Pro, HumanEval, and GPQA Diamond, outperforming closed models several times its size.

What makes it stand out: it's a unified model that handles fast chat, deep chain-of-thought reasoning, vision input, and agentic coding within a single architecture — no separate variants. It runs on a single A100 GPU, or on consumer hardware with quantization, making on-premise and privacy-preserving deployments genuinely viable without a massive infrastructure budget.

Key features: - Apache 2.0 license: commercially deploy and fine-tune freely - Multimodal: handles text, images, and structured documents - Agentic coding: integrates Devstral-style tool use and task completion - Ecosystem support: available on vLLM, llama.cpp, SGLang, and Hugging Face Transformers - Runs on consumer hardware via quantization

Mistral Small 4 is ideal for developers building commercial products who need a capable, self-hostable foundation model without the licensing overhead of proprietary APIs. It's particularly strong as a base for fine-tuning on domain-specific tasks, running local coding agents, or powering privacy-first enterprise deployments.

Available now on Hugging Face, Mistral's platform (La Plateforme), and inference frameworks like vLLM. API access starts on competitive pay-per-token pricing.