KAT-Coder

Today, we're thrilled to announce two groundbreaking models in our KAT series: KAT-Dev-32B and KAT-Coder — representing accessible excellence and ultimate performance in code intelligence. We are excited to introduce KAT-Dev-32B - our new open-source 32B-parameter model for software engineering tasks. We have released KAT-Dev-32B to the community for further research and development and you can find it at https://huggingface.co/Kwaipilot/KAT-Dev. Moreover, KAT-Coder is our most powerful variant.

Key Contributions

Our KAT-Coder and KAT-Dev-32B are optimized via several stages of training, including a mid-training stage, supervised fine-tuning (SFT) & reinforcement fine-tuning (RFT) stage and an large-scale agentic reinforcement learning (RL) stage. In summary, our contributions include:

🎯 Mid-Training: We observe that adding extensive training for tool-use capability, multi-turn interaction, and instruction-following at this stage may not yield large performance gains in the current results (e.g., on leaderboards like SWE-bench), but it will have a significant impact on the subsequent SFT and RL stages.

🎯 SFT & RFT: We meticulously curated eight task types and eight programming scenarios during the SFT stage to ensure the model's generalization and comprehensive capabilities. Moreover, before RL, we innovatively introduced an RFT stage with "teacher trajectories" annotated by human engineers as guidance during training.

🎯 Agentic RL Scaling: Scaling agentic RL hinges on three challenges: efficient learning over nonlinear trajectory histories, leveraging intrinsic model signals, and building scalable high-throughput infrastructure. We address these with prefix caching on logprob computation, entropy-based trajectory pruning, and SeamlessFlow architecture.

Moreover, to achieve an enhanced performance on SWE-Bench, we further equip our KAT-Coder with an extra reflection step and iteratively refine the generated outputs.

You can learn more details about our models in https://kwaipilot.github.io/KAT-Coder/ and how we perform reflection in https://github.com/kwaipilot/KAT-Coder-Agent/blob/main/KAT-Coder_Report.md.

Performance

This pull request includes the performance of the KAT-Coder on the SWE-bench verify.

Submission summary for 20250916_KAT-Coder on SWE-bench verified split
==================================================
Resolved 382 instances (76.4%)
==================================================
Resolved by Repository
- astropy/astropy: 13/22 (59.09%)
- django/django: 184/231 (79.65%)
- matplotlib/matplotlib: 24/34 (70.59%)
- mwaskom/seaborn: 1/2 (50.0%)
- pallets/flask: 1/1 (100.0%)
- psf/requests: 2/8 (25.0%)
- pydata/xarray: 18/22 (81.82%)
- pylint-dev/pylint: 6/10 (60.0%)
- pytest-dev/pytest: 15/19 (78.95%)
- scikit-learn/scikit-learn: 30/32 (93.75%)
- sphinx-doc/sphinx: 28/44 (63.64%)
- sympy/sympy: 60/75 (80.0%)
==================================================
Resolved by Time
- 2013: 1/3 (33.33%)
- 2014: 0/2 (0.0%)
- 2015: 0/1 (0.0%)
- 2016: 2/2 (100.0%)
- 2017: 14/16 (87.5%)
- 2018: 19/24 (79.17%)
- 2019: 78/98 (79.59%)
- 2020: 89/108 (82.41%)
- 2021: 58/86 (67.44%)
- 2022: 78/102 (76.47%)
- 2023: 43/58 (74.14%)

Checklist

• Is a pass@1 submission (does not attempt the same task instance more than once)
• Does not use SWE-bench test knowledge (PASS_TO_PASS, FAIL_TO_PASS)
• Does not use the hints field in SWE-bench
• Does not have web-browsing OR has taken steps to prevent lookup of SWE-bench solutions via web-browsing