Mistral Small 4

Mistral Small 4 (March 16, 2026) is Mistral's unified mid-tier model: a 119B-parameter mixture-of-experts (128 experts, ~6B active per token) that folds the previously separate instruct, Magistral reasoning, Pixtral vision and Devstral coding lines into a single checkpoint with a per-request reasoning_effort dial. It ships under Apache 2.0 with weights on Hugging Face (mistralai/Mistral-Small-4-119B-2603), a 256K context window, and day-one vLLM/SGLang/NVIDIA NIM support.

The catch is memory, not compute: MoE sparsity makes inference cheap per token, but all 119B parameters must sit in memory — even aggressive 3-bit quants need ~54GB, so true local use means multi-GPU workstations, not laptops. Spatial reasoning benchmarks also show gaps versus dense flagships. Still, for teams consolidating several model endpoints into one permissively licensed open model, it is one of the strongest 2026 options.