Documented coding focus
According to the model card, KAT-Dev-72B-Exp is an open-source 72B-parameter model for software engineering tasks and the experimental reinforcement-learning version of KAT-Coder.
Open Source Model Profile · Kwaipilot
KAT-Dev-72B-Exp
KAT-Dev-72B-Exp is a 72.71B-parameter Qwen2 software-engineering release from Kwaipilot. According to the model card, it is the experimental RL version of KAT-Coder with a reported SWE-Bench figure.
- Publisher
- Kwaipilot
- Task
- text-generation
- Model type
- qwen2
- License
- apache-2.0
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
KAT-Dev-72B-Exp is published by Kwaipilot as a qwen2-based text-generation release. The captured configuration identifies Qwen2ForCausalLM and Safetensors metadata reports 72706203648 parameters. According to the model card, it is an open-source 72B-parameter software-engineering model and the experimental reinforcement-learning version of KAT-Coder.
Recorded capabilities
Publisher-reported SWE-Bench figure
According to the model card, the model achieves 74.6% accuracy on SWE-Bench Verified with the SWE-agent scaffold, with evaluation parameters including temperature 0.6 and 150 maximum turns.
Documented RL training work
According to the model card, training work included a rewritten attention kernel with a redesigned engine for shared prefix trajectories, plus advantage-distribution reshaping based on pass rates.
Transformers usage record
According to the model card, usage loads the tokenizer and causal-LM model with automatic dtype and device mapping, then applies the chat template before generation.
Qwen2 72.71B Transformers record
Captured config identifies Qwen2ForCausalLM and qwen2, with Safetensors metadata reporting 72706203648 parameters and transformers library support.
Use cases in the source record
- Software-engineering text generation using the captured transformers Qwen2 configuration with the documented chat-template workflow.
- SWE-agent scaffold experiments referencing the publisher-reported 74.6% SWE-Bench Verified setup with temperature 0.6 and 150 maximum turns.
Limitations and unknowns
- No context-window value was extracted from this record; the card example shows max_new_tokens 65536, which is a generation setting rather than a confirmed model limit.
- Provider state is historical snapshot data, not independently refreshed current availability.
- The 74.6% SWE-Bench Verified figure is a publisher-reported result tied to the SWE-agent scaffold and was not independently measured by Ethen.
- Release-announcement characterizations such as latest and most powerful are publisher claims and were not independently verified by Ethen.
Source and provenance
Source: Kwaipilot/KAT-Dev-72B-Exp
Captured: Unknown. Processed: 2026-09-07T19:34:33.352233+00:00.
News 🔥 We’re thrilled to announce the release of KAT-Dev-72B-Exp , our latest and most powerful model yet! 🔥 You can now try our strongest proprietary coder model KAT-Coder directly on the StreamLake platform for free . Highlights KAT-Dev-72B-Exp is an open-source 72B-parameter model for software engineering tasks. On SWE-Bench Verified, KAT-Dev-72B-Exp achieves 74.6% accuracy ⚡ — when evaluated strictly with the SWE-agent scaffold . KAT-Dev-72B-Exp is the experimental reinforcement-learning version of the KAT-Coder model. Through this open-source release, we aim to reveal the technical innovations behind KAT-Coder’s large-scale R…
F001F002F003F004F005F006F007F008F009F010F011F013F014F015F016F017F018F019F020