Devstral 2 123B Instruct 2512
mistralai · released 2025-12-09 · other license
Devstral 2 (123B) Instruct is Mistral's Dec 2025 open-weight coding model (modified MIT). It reaches 72.2% SWE-bench Verified and 40.5% Terminal-Bench (vendor), one of the strongest open-weight code agents.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 125.03B total |
| Architecture | ministral3 |
| Context window | 262K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | 72.2% |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 7.22 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 74 GB | 72.5 GB | — | 262K |
Strengths & weaknesses
Strengths: SOTA open-weight coding agent: 72.2% SWE-bench Verified at 123B; 256K context, strong multi-file/repo agentic coding; Cost-efficient, permissive modified MIT
Weaknesses: Still trails leading closed models in human-eval coding preference; Needs 4+ H100-class GPUs