Mistral Small 3
Last updated: April 22, 2026
Specs
| Vendor | Mistral AI |
|---|
| Released | January 30, 2025 |
|---|
| Context window | 128K tokens |
|---|
| Input | $0.2 / M tokens |
|---|
| Output | $0.6 / M tokens |
|---|
Benchmarks
| mmlu | 81.2 |
|---|
| gpqa | 45.3 |
|---|
| humanEval | 84.8 |
|---|
| arenaElo | 1266 |
|---|
| sweBench | 30 |
|---|
| sweBenchPro | 20 |
|---|
| liveCodeBench | 45 |
|---|
| aiderPolyglot | 42 |
|---|
| terminalBench | 16 |
|---|
| mmluPro | 62 |
|---|
| hle | 4 |
|---|
| aime | 45 |
|---|
| math500 | 75 |
|---|
Strengths
- Open weights under Apache 2.0
- Very fast on modest hardware
- EU-based vendor, GDPR-friendly by default
Weaknesses
- Not frontier-class on reasoning
- No native multimodal
- Smaller context than the 256K+ crowd
Best for
- On-premise European deployments
- Classification and routing at high volume
- Edge inference where latency matters
Compare with other AI models