by Qwen
KAT-Dev (32B) is an open-source 32B parameter model specifically designed for software engineering tasks. It achieved a 62.4% resolution rate on the SWE-Bench Verified benchmark, ranking fifth among all open-source models of various scales. The model is optimized through multiple stages, including intermediate training, supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), as well as large-scale agent reinforcement learning (RL). Based on Qwen3-32B, its training process lays the foundation for subsequent fine-tuning and reinforcement learning stages by enhancing fundamental abilities such as tool usage, multi-turn interaction, and instruction following. During the fine-tuning phase, the model not only learns eight carefully curated task types and programming scenarios but also innovatively introduces a reinforcement fine-tuning (RFT) stage guided by human engineer-annotated “teacher trajectories.” The final agent reinforcement learning phase addresses scalability challenges through multi-level prefix caching, entropy-based trajectory pruning, and efficient architecture.
KAT-Dev (32B) is an open-source 32B parameter model specifically designed for software engineering tasks. It achieved a 62.4% resolution rate on the SWE-Bench Verified benchmark, ranking fifth among all open-source models of various scales. The model is optimized through multiple stages, including intermediate training, supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), as well as large-scale agent reinforcement learning (RL). Based on Qwen3-32B, its training process lays the foundation for subsequent fine-tuning and reinforcement learning stages by enhancing fundamental abilities such as tool usage, multi-turn interaction, and instruction following. During the fine-tuning phase, the model not only learns eight carefully curated task types and programming scenarios but also innovatively introduces a reinforcement fine-tuning (RFT) stage guided by human engineer-annotated “teacher trajectories.” The final agent reinforcement learning phase addresses scalability challenges through multi-level prefix caching, entropy-based trajectory pruning, and efficient architecture.
kat-dev has a 128,000 token context window.
On AIHubMix, kat-dev costs $0.14 per million input tokens and $0.55 per million output tokens.
kat-dev accepts text input.
kat-dev supports tool calling. Per-protocol parameter support is listed in the capability table on this page.
kat-dev is available through the AIHubMix unified API. The API is OpenAI-compatible: point your OpenAI SDK at https://aihubmix.com/v1, use your AIHubMix API key, and set the model name to kat-dev — no other code changes needed.
kat-dev is developed by Qwen. AIHubMix aggregates it alongside models from other providers behind one API and one bill.
Qwen 3.8 Max Preview(Qwen3.8-Max-Preview) is the latest-generation foundation model in…
qwen-audio-3.0-tts-flash is a high-performance speech synthesis large model optimized for…
qwen-audio-3.0-tts-plus is a high-performance speech synthesis large model designed for…
HappyHorse-1.1-I2V supports image-to-video generation, further enhancing visual texture…
HappyHorse-1.1-R2V supports reference-based video generation, further improving the…
HappyHorse-1.1-T2V supports text-to-video generation, further enhancing text semantic…
Use kat-dev via the AIHubMix unified API — one interface for every major LLM.