Mistral Small 24B Instruct 2501
mistralai · released 2025-01-30 · apache-2.0 license
Mistral Small 24B Instruct 2501 (Jan 2025, Apache-2.0) is a knowledge-dense 24B rivaling much larger models: vendor MMLU-Pro 66.3, IFEval 82.9.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 23.57B total |
| Architecture | mistral |
| Context window | 33K tokens |
| Knowledge cutoff | 2023-10-31 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | 66.3% |
| BFCL v3 (tool use) | — |
| Composite score | 7.31 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 15.2 GB | 13.7 GB | — | 33K |
Strengths & weaknesses
Strengths: SOTA for its size class (<70B), rivaling Llama-3.3-70B / Qwen2.5-32B; Strong function-calling and JSON output; Runs on a single 4090; Apache-2.0
Weaknesses: 32K context smaller than many competitors; No built-in safety guardrails