KAT-Coder-V2.5-Dev (35B-A3B)
Open-weight coding model: 35B MoE, 3B active, 262K native context, Apache-2.0. Language-model weights only (no vision tower). Qwen-style reasoning and tool parsers.
Recommended quants
- Q4_K_M (21.4 GB)
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF - Q3_K_XL (17.3 GB) / IQ2_XXS (9.8 GB)
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUFLower memory, lower quality. - Q8_0 (36.9 GB)
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF - APEX GGUF
mudler/KAT-Coder-V2.5-Dev-APEX-GGUFMost-downloaded community quant; sizes not checked.
Chat template Works
Qwen3-style template; thinking toggled with enable_thinking.
Fix: vLLM: --language-model-only is REQUIRED (no vision weights shipped). SGLang may fail at startup building vision components.
Sampling settings
| Thinking (model card example) | temp 1.0, top_p 0.95, top_k 20, presence_penalty 1.5 |
|---|---|
| Non-thinking (model card example) | temp 0.7, top_p 0.8, top_k 20, presence_penalty 1.5 |
Runtimes
| vLLM | Caveats | Needs --language-model-only; --reasoning-parser qwen3 --tool-call-parser qwen3_coder. |
|---|---|---|
| llama.cpp | Works | bartowski GGUFs. |
| Ollama | Unverified | No official library page found. |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | Vendor benchmarks used claude_code as the agent harness. |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
- https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev
- https://huggingface.co/bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)