Kwaipilot's KAT-Coder-V2.5 Hits 69.4% on Bug Fixing, Now Runs on Consumer GPUs
APEX quantization compresses Kwaipilot's 35B agentic coding MoE into GGUF files that run locally on consumer GPUs while preserving benchmark accuracy.
Takeaways
- APEX GGUF quantizations of Kwaipilot's KAT-Coder-V2.5-Dev, a 35B MoE agentic coding model, released Apache-2.0.
- Base model scores 69.4% on SWE-bench Verified with only 3B active parameters per token.
- Seven quantization profiles from Mini to I-Quality, targeting consumer GPUs up to workstation cards.
- Role-aware quantization keeps attention and shared experts precise while compressing middle-layer routed experts.
- Runs out of the box with llama.cpp, Ollama, LM Studio, LocalAI, and any OpenAI-compatible agent tooling.
- Text-only release; vision components stripped, so vLLM needs the language-model-only flag.
KAT-Coder-V2.5-Dev-APEX-GGUF packages Kwaipilot's Mixture-of-Experts coding model into a set of quantized GGUF files designed to run on hardware most developers actually own. The quantization uses a technique called APEX, which preserves coding accuracy on MoE architectures while cutting the memory footprint enough to fit on a single consumer GPU.
The base model is KAT-Coder-V2.5-Dev, a 35B-parameter MoE built by Kwaipilot for autonomous software engineering inside executable repositories. With 3B parameters activated per token, it scores 69.4% on SWE-bench Verified, placing it ahead of many larger open models on repo-level bug fixing. Its agentic training lets it execute commands, read repository files, and iterate on tasks inside real sandboxed environments rather than generating code in isolation.
Why APEX handles MoE better than uniform quantization
Standard uniform quantization degrades MoE models because expert tensors and attention tensors have very different numerical sensitivities. APEX classifies tensors by function and applies a layer-wise precision gradient: the first and last few layers retain higher precision while middle layers compress harder. I-variants add imatrix calibration using a diverse corpus of chat, code, reasoning, tool-calling, and agentic traces.