Qwen3.5-9B SWE-GRPO
a coding and software engineering model
1 min readupdated
This is a fine-tuned version of the base Qwen/Qwen3.5-9B (9.65B parameters) trained with GRPO (a reinforcement learning method based on group relative policy optimization) to act as a software engineering agent. Trained on 1,000 tasks from the swe-rl-curated-1k dataset using sandbox test-execution rewards, it outperforms the base model at code editing and bug fixing while significantly cutting down on output formatting errors. It is a great fit for developers and researchers building autonomous AI agents to write and debug code.
Key features
- Trained on 1,000 tasks using executable test-based rewards
- Training set pass rate jumped from 18% to ~70%
- Stronger performance on SWE-bench Verified (up to 60.6% on a task subset)
- Major drop in syntax errors and malformed tool calls
What you can do
- Plug the model into an agentic loop to automatically fix bugs in repositories
- Run it inside automated Python testing and development pipelines
- Handle end-to-end coding tasks autonomously from text descriptions
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("mliu444/qwen3.5-9b-swe-grpo", revision="iter59", dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("mliu444/qwen3.5-9b-swe-grpo", revision="iter59")
Read more
DetailsBenchmarks: 60.6% on 500 SWE-bench Verified tasks (step 59)
Training: GRPO, constant learning rate 1e-6, Megatron trainer
Limitations: trained on 1,000 Python repositories; evaluated only within a specific agent environment










Comments