Mistral AI · open weights · updated July 2026

Mistral Small 4

Parameters
119B / 6.5B active
Context window
256K tokens
License
Apache 2.0
Modalities
text, image
Released
2026-03-01

Mistral Small 4 is the European entry in the 2026 open-weight field and arguably the best production-per-watt model available: 119B total parameters with only 6.5B active, 256K context, multimodal input, a reasoning mode, native function calling and JSON output — all under Apache 2.0.

Mistral's positioning is deployment-first rather than benchmark-first: Small 4 targets the regulated-industry and on-premise market where EU jurisdiction, clean licensing, and single-GPU serving matter more than leaderboard positions. For European enterprises with data-sovereignty requirements it is frequently the default choice.

Mistral Small 4 specifications

VendorMistral AI (France)
ArchitectureSparse Mixture-of-Experts, multimodal input
Total parameters119B
Active parameters6.5B
Context window256K tokens
Modalitiestext, image
LicenseApache 2.0 — permissive, patent-granting, free for commercial use
Release date2026-03-01
WeightsHugging Face · GitHub

Running Mistral Small 4 locally

119B total / 6.5B active: about 65–70GB at 4-bit — a single 80GB GPU, dual 48GB cards, or a 96GB+ unified-memory Mac. The strongest production-friendly European open model to self-host on modest hardware.

What Mistral Small 4 is best for

  • EU data-sovereignty and regulated-industry deployments
  • Single-GPU production serving (6.5B active parameters)
  • Function calling and structured-output pipelines

Frequently asked questions

Why choose Mistral Small 4 over larger Chinese or US models?

Jurisdiction and deployability. It is the strongest open-weight model from an EU vendor under Apache 2.0, and its 6.5B-active MoE design serves on a single 80GB GPU — a combination that regulated European enterprises often cannot get from trillion-parameter alternatives.