Mistral Small 4
Released 16 March 2026 — one model replacing three. Reasoning, vision and coding unified in a 119B-parameter MoE that activates just 6B parameters per token.
Mistral Small 4 specs at a glance
| Maker | Mistral AI |
| Released | 16 March 2026 |
| Architecture | Mixture-of-Experts · 119B total params · 128 experts · 4 active per token |
| Active parameters | ~6B per token |
| Capabilities | Reasoning + multimodal vision + agentic coding, unified |
| API pricing | TBD — not yet verified; see Mistral's pricing page |
| Context window | TBD — not yet verified |
| Best at | Efficient European-hosted deployments needing one general model |
Calculate LLM API costs → This model has no confirmed USD per-token rate card, so it is not in the calculator — compare 16 published rate cards instead.
What's new in Mistral Small 4
Mistral's 2025 lineup asked you to choose a specialist: Magistral for reasoning, Pixtral for images, Devstral for agentic coding. Small 4 collapses that choice — one model handles all three, which simplifies deployment and routing for teams that were juggling per-task endpoints.
The efficiency story is the MoE ratio: 119B total parameters, but only ~6B active per token (4 of 128 experts). That keeps serving costs closer to a small dense model while retaining large-model breadth.
Where it sits in the 2026 Mistral lineup
| Model | Released | Size | Role |
|---|---|---|---|
| Mistral Large 3 | 2 Dec 2025 | 675B total / 41B active | Flagship — largest open-weight MoE from a major lab |
| Mistral Medium 3.5 | 29 Apr 2026 | — | Enterprise balance point |
| Mistral Small 4 | 16 Mar 2026 | 119B total / 6B active | Efficient unified tier |
| Voxtral TTS | 23 Mar 2026 | 3B-class | First Mistral audio/TTS model |
Who should use Mistral Small 4
- Teams consolidating multi-model stacks — one endpoint instead of three specialists.
- Cost-efficient serving — the 6B-active MoE keeps per-token compute low.
- European / data-residency requirements — Mistral remains the leading EU-based frontier lab.
How Mistral Small 4 compares
- Gemma 4 — Google's open efficient family →
- DeepSeek V4 — the open-weight heavyweight →
- Claude Haiku 4.5 — the closed efficient tier →
Frequently asked questions
When was Mistral Small 4 released?
16 March 2026.
What makes it different?
It unifies three previously separate Mistral models — Magistral (reasoning), Pixtral (vision) and Devstral (agentic coding) — into a single system.
What architecture does it use?
Mixture-of-Experts: 119B total parameters, 128 experts, ~6B parameters (4 experts) active per token.
How much does it cost?
We haven't verified official per-token API pricing yet — check Mistral's pricing page for current rates.
How does it relate to Mistral Large 3?
Large 3 (Dec 2025, 675B/41B active) remains the flagship; Small 4 is the efficient unified tier below it, with Medium 3.5 in between.