35B total / 3B active (MoE; 34.66B checkpoint parameters)
262,144 tokens
Apache 2.0
Open-weight developer-oriented member of the KAT-Coder-V2.5 family, post-trained specifically for autonomous repository-level software engineering, agentic coding, tool use, and long-horizon debugging workflows. KAT-Coder-V2.5-Dev is built by post-training Qwen3.6-35B-A3B and retains its hybrid Qwen3.5-MoE architecture identifier. The checkpoint contains approximately 34.66B BF16 parameters, marketed as 35B total with 3B activated per token. Its 40-layer language backbone alternates three GatedDeltaNet linear-attention layers with one full-attention layer, yielding 30 linear-attention and 10 full-attention layers. It uses a 2,048-dimensional hidden state, 16 attention heads, two key-value heads, 256 routed experts with eight selected per token, one shared expert, and a 262,144-token native context window. Although the inherited configuration declares image and video components, Kwaipilot explicitly states that the released checkpoint contains only language-model weights and must be served in text-only mode; no vision tower or multimodal capability is included. The model supports both default thinking behavior and a lower-latency non-thinking mode, structured tool calling, and optional preservation of reasoning traces across turns for multi-step agent workflows. Post-training begins with supervised fine-tuning on 127K examples and continues with ten stable reinforcement-learning epochs. The RL pipeline retains the broader KAT-V2.5 framework's token-in/token-out consistency, truncated importance sampling, reliability-hardened sandboxes and verifiers, and hierarchical rewards derived from execution feedback, while adding penalties for excessive parallel tool calls, failed or empty tool calls, and repeated content. Kwaipilot reports that this reduced malformed tool labels from 9.34% to 0.28% and eliminated measured single-turn continuous repetition from 0.34% to zero. Under its standardized in-house evaluation harness, the team reports scores of 69.40 on SWE-bench Verified, 63.00 on SWE-bench Multilingual, 45.96 on SWE-bench Pro, 41.02 averaged across two Terminal-Bench 2.1 harnesses, 93.43 on PinchBench, 44.20 on SciCode, and 46.21 on KAT-Code-Bench. The official artifacts support deployment through Transformers, vLLM, SGLang, and KTransformers, with vLLM requiring its language-model-only option because the vision weights are absent.