nebius/SWE-rebench-openhands-Qwen3-30B-A3B

Model Summary

SWE-rebench-openhands-Qwen3-30B-A3B is a 30B Rejection Sampling Fine-Tuning (RFT) checkpoint derived from Qwen/Qwen3-30B-A3B-Instruct-2507, trained on the newly released nebius/SWE-rebench-openhands-trajectories dataset. Training used a maximum sequence length of 131k tokens.

Model	Size	Maximum Number of Turns = 100				Maximum Number of Turns = 500
		Pass@1	Pass@5	Pass@1	Pass@5	Pass@1	Pass@5	Pass@1	Pass@5
30B scale
Qwen3-30B-A3B-Instruct-2507	30B	25.2	44.8	11.8	24.4	25.7	44.2	14.2	26.5
Qwen3-Coder-30B-A3B-Instruct	30B	51.9	67.3	28.7	42.8	50.0	63.0	28.1	38.7
nebius/SWE-rebench-openhands-Qwen3-30B-A3B (Ours)	30B	49.7 (+24.5)	65.4 (+20.6)	28.1 (+16.3)	38.7 (+14.3)	50.3 (+24.6)	68.3 (+24.1)	28.1 (+13.9)	38.7 (+12.2)
100B+ scale
GLM-4.5-Air	106B	58.2	73.5	33.8	42.8	-	-	-	-
200B+ scale
Qwen3-235B-A22B-Instruct-2507	235B	45.2	65.9	29.3	44.8	46.2	67.5	25.3	40.8
nebius/SWE-rebench-openhands-Qwen3-235B-A22B (Ours)	235B	59.9 (+14.7)	73.9 (+8.0)	35.1 (+5.8)	46.9 (+2.1)	61.7 (+15.5)	74.3 (+6.8)	34.2 (+8.9)	44.8 (+4.0)
300B+ scale
GLM-4.5	355B	64.4	76.2	33.8	44.8	-	-	-	-
Qwen3-Coder-480B-A35B-Instruct	480B	64.7	75.8	36.3	44.8	66.5	77.8	35.5	42.8

Table 1. Pass@1 (averaged over 5 runs) and Pass@5 for OpenHands agent with the maximum number of turns set to 100 (highlighted in yellow) and 500 (highlighted in green). Metrics are reported in percentages. Deltas vs base models are shown in parentheses for fine-tuned models.

We explicitly excluded all SWE-bench Verified and SWE-rebench September issues from training to avoid contamination. SWE-rebench Verified was additionally decontaminated on repository level.

When evaluated with the OpenHands (v0.54.0) agent, our 30B model:

• Substantially improves over the base Qwen3-30B-A3B-Instruct-2507 model, with +24.6 Pass@1 and +24.1 Pass@5 gains at 500-turn settings.
• Matches or surpasses the specialized Qwen3-Coder-30B-A3B-Instruct baseline at 500 turns on SWE-bench Verified (50.3% vs 50.0% Pass@1; 68.3% vs 63.0% Pass@5).
• Generalizes to longer interaction horizons, despite training on trajectories capped at 100 turns.

For more details see our report in Nebius blog.

---

Best Practices

1. Deployment:

   • Use the following configuration to serve the model with vLLM:

     VLLM_USE_V1=1 vllm serve nebius/SWE-rebench-openhands-Qwen3-30B-A3B
       --tensor-parallel-size 8
       --served-model-name qwen_3_instruct_2507
       --disable-log-requests
       --enable-prefix-caching
       --max-model-len 131072
       --enable-auto-tool-choice
       --tool-call-parser hermes
     

     Tested using vllm/vllm-openai:v0.9.0 Docker image.

2. Sampling Parameters:

   • For optimal performance, we recommend Temperature=0.7, TopP=0.8, TopK=20, and MinP=0 that are consistent with the base model.

---

Citation

@article{trofimova2025openhandstrajs,
  title={OpenHands Trajectories with Qwen3-Coder-480B-A35B-Instruct},
  author={Trofimova, Maria and Shevtsov, Anton and Ibragim, Badertdinov and Pyaev, Konstantin and Karasik, Simon and Golubev, Alexander},
  year={2025},
  journal={Nebius blog},
  note={}
}