Back to the index
Mistral AI · Mistral
No. 22

Mistral Small 4

Tiny Apache-2.0 multimodal model unifying reasoning, vision and coding.

FastOpen weightsText, Image
Field notes
Context window
262Ktokens
Max output
33Ktokens
Input price
$0.15/ 1M tokens
Output price
$0.60/ 1M tokens
Modalities
Text, Image
Knowledge cutoff
—
Released
Mar 2026
Parameters
24B
License
Apache 2.0
Approx. speed
~169 tok/s
Overview

What this model is

Mistral Small 4 (2603) is a compact Apache-2.0 model that merges Mistral's previously separate Magistral (reasoning), Pixtral (vision) and Devstral (coding) lines into one small, cheap set of weights with an optional reasoning mode. It is one of the best-value small multimodal models and fully unrestricted for commercial use, but it is not built for hard, long agentic coding. Ideal for edge, local, and high-volume deployments where openness and cost dominate.

Strengths

  • Fully open Apache 2.0 license
  • Very fast output (~169 t/s) and cheap
  • Unified reasoning + vision + coding
  • 256K context, self-hostable on modest hardware

Trade-offs

  • Absolute capability below larger models
  • Agentic reliability limited
  • Weaker on hardest reasoning/coding
  • Small-model recall limits on long context

Best for

  • Local and edge deployment
  • High-volume cheap inference
  • Fine-tuning with no license friction
  • Lightweight multimodal tasks

Not ideal for

  • Complex agentic coding
  • Frontier reasoning workloads
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning66Coding64Math64Writing72Knowledge72Speed90Agentic58Vision66Multilingual80Long Context76
Reasoning
66
Coding
64
Math
64
Writing
72
Knowledge
72
Speed
90
Agentic
58
Vision
66
Multilingual
80
Long Context
76
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
21/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
21%
τ-benchindependent
Agenticτ-bench
65.8%
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.

Llama 4 ScoutMeta AI

Ultra-long 10M-token context, efficient single-GPU open model.

Qwen3-30B-A3B-Instruct-2507Alibaba Qwen

Small, fast open-weight MoE with big-model manners.

Magistral Medium 1.2Mistral AI

Mistral's transparent-chain reasoning flagship, now multimodal.