SWE-bench Verified Leaderboard
SWE-bench Verified tests whether a model can resolve real GitHub issues from popular open-source Python projects, scored on the official swebench.com leaderboard as the percentage of human-validated issues actually fixed. It is the headline measure of practical, agentic software-engineering ability — where open-weight models like Qwen3-Coder, GLM-4.6, Kimi K2 and DeepSWE are now competitive with the frontier.
Source: swebench20 open models ranked+142 proprietaryData through Feb 2026
Open models ranked on SWE-bench Verified
# shows rank among open models / rank overall (including proprietary).
| # | Model | Score | VRAM @ Q4 |
|---|---|---|---|
| 1 / 10 | MiniMax M2.5 · 228.7B | 75.8% | ~138 GB |
| 2 / 25 | GLM 5 · 753.9B | 72.8% | ~457 GB |
| 3 / 33 | Kimi K2 Instruct 0905 · 1026.5B | 71.2% | ~620 GB |
| 4 / 36 | Kimi K2.5 · 1026.9B | 70.8% | ~620 GB |
| 5 / 44 | DeepSeek V3.2 · 685.4B | 70.0% | ~415 GB |
| 6 / 47 | Qwen3 Coder 480B A35B Instruct · 480.2B | 69.6% | ~289 GB |
| 7 / 49 | GLM 4.6 · 356.8B | 68.2% | ~215 GB |
| 8 / 60 | Kimi K2 Instruct · 1026.4B | 65.4% | ~620 GB |
| 9 / 65 | GLM 4.5 · 358.3B | 64.2% | ~216 GB |
| 10 / 67 | Kimi K2 Thinking · 1026.4B | 63.4% | ~620 GB |
| 11 / 73 | MiniMax M2 · 228.7B | 61.0% | ~138 GB |
| 12 / 75 | Qwen3 Coder 30B A3B Instruct · 30.5B | 60.4% | ~19 GB |
| 13 / 78 | DeepSWE Preview · 32.8B | 58.8% | ~20 GB |
| 14 / 84 | Devstral Small 2 24B Instruct 2512 · 24.0B | 56.4% | ~15 GB |
| 15 / 88 | Devstral 2 123B Instruct 2512 · 125.0B | 53.8% | ~76 GB |
| 16 / 104 | Qwen2.5 Coder 32B Instruct · 32.8B | 47.0% | ~21 GB |
| 17 / 105 | Devstral Small 2505 · 23.6B | 46.8% | ~15 GB |
| 18 / 116 | DeepSeek v3 0324 · 684.5B | 42.0% | ~415 GB |
| 19 / 130 | Devstral Small 2507 · 23.6B | 38.0% | ~15 GB |
| 20 / 142 | GPT OSS 120B · 116.8B | 26.0% | ~71 GB |
Score vs model size
Which models give the most quality for their size — the ones worth running locally.
- Devstral Small 2505, 24B, score 46.8% — on the efficiency frontier (best score at its size or smaller).
- Devstral Small 2 24B Instruct 2512, 24B, score 56.4% — on the efficiency frontier (best score at its size or smaller).
- Qwen3 Coder 30B A3B Instruct, 31B, score 60.4% — on the efficiency frontier (best score at its size or smaller).
- MiniMax M2, 229B, score 61.0% — on the efficiency frontier (best score at its size or smaller).
- MiniMax M2.5, 229B, score 75.8% — on the efficiency frontier (best score at its size or smaller).
SWE-bench Verified: frequently asked questions
- What is the best open LLM on SWE-bench Verified?
- MiniMax M2.5 is the top open model on SWE-bench Verified, scoring 75.8%. Among all models tested — including proprietary ones — it ranks #9. The top model overall is Sonar Foundation Agent + Claude 4.5 Opus at 79.2%.
- What's the best SWE-bench Verified model you can run on a 24 GB GPU?
- Qwen3 Coder 30B A3B Instruct is the highest-scoring open model that fits in 24 GB at 4-bit quantization (~19 GB), scoring 60.4% on SWE-bench Verified.
- Can open models match proprietary models on SWE-bench Verified?
- Not quite on SWE-bench Verified: the strongest proprietary model (Sonar Foundation Agent + Claude 4.5 Opus) scores 79.2%, ahead of the best open model (MiniMax M2.5) at 75.8% — but you can run the open one yourself.
Scores aggregated from swebench. llmrun does not run this benchmark — see the source for methodology, or the about benchmarks for what it measures.