Mistral Small 3
Mistral Small 3, a 24B Apache 2.0 model, matches Llama 3.3 70B on instruction tasks at over 3x the speed.
Latency-optimized 24B with fewer layers, at over 81% MMLU and about 150 tokens/s. Mistral committed to Apache 2.0 for its general-purpose models, after Large 2, Codestral and Pixtral Large shipped under restricted licenses.
- Date
- Thursday, 30 January 2025
- Lab
- Mistral AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| MMLU | over 81% | company |
| Throughput | ~150 tokens/s vs Llama 3.3 70B more than 3x faster on same hardware | company |
Sources
This record was checked against its sources on 6 October 2026. How we check