LocalClaw/LLM/KAT-Coder V2.5 Dev
Open-weight MoE

KAT-Coder V2.5 Dev

Catalogue summary: Kwaipilot open-weight Apache 2.0 coding-agent MoE built from Qwen3.6-35B-A3B. Text-only release with 262K context, 35B total / 3B active parameters and established GGUF paths from bartowski and LocalAI/APEX for LM Studio, llama.cpp and LocalAI testing on 48GB+ machines.

Repository editorial metadata; verify comparative claims in the linked upstream material.

48 GB catalogue minimum48 GB RAMQ4_K_MCoding assistant
Choose an app

Pick the app you already use. No terminal and no command to copy.

Compare models
Open in LM StudioOpens the app on this model's download screenOpen in UnslothLaunches Unsloth Desktop on this model's download page
OllamaNot available for this model
Open on Hugging FaceFiles, licence and available downloadsOpen llama.cpp setupAdvanced option: opens the model filesUse with LocalClawOptional workspace after the model is installed

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

Parameters
35B (3B active, MoE)
Minimum RAM
48 GB
Model size
21.39 GB
Quantization
Q4_K_M

Can KAT-Coder V2.5 Dev run locally?

KAT-Coder V2.5 Dev has a catalogue minimum of 48 GB RAM with Q4_K_M. Actual memory use and speed vary by context length, runtime, backend and system headroom.

Use kat-coder-v2.5-dev as the catalogue search term in a compatible runtime, and confirm the available format on the upstream repository before download.

Model sourcePublic GGUF repository (verified 2026-09-07): bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUFUpstream releasePaper or model cardLicence text
codereasoningagenticlong-contexttool-callingpower

Install path

01
Check RAM fitMinimum 48 GB RAM. Start with the Q4_K_M quant.
02
Load the modelSearch kat-coder-v2.5-dev in LM Studio.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Catalogue record

  • Family: kat-coder
  • Parameters: 35B (3B active, MoE)
  • Recommended quantization: Q4_K_M
  • Catalogue minimum RAM: 48 GB
  • Catalogue model size: 21.39 GB
  • Tags: code, reasoning, agentic, long-context, tool-calling, power

Practical limits

  • Catalogue RAM is a minimum estimate, not a guarantee for every context length or runtime.
  • Speed and memory use vary by quantization, backend, context length and system headroom.
  • Verify architecture, licence and usage restrictions in the linked upstream material before deployment.

Catalogue tags

  • code
  • reasoning
  • agentic
  • long-context
  • tool-calling
  • power

Capability profile

Repository catalogue ratings used by LocalClaw's editorial rubric. They are not a standardized third-party benchmark.

speed
6
quality
8
coding
10
reasoning
8

Technical notes

Developer
Kwaipilot
License
Apache 2.0
Context window
262,144 tokens
Architecture
Qwen3.6-35B-A3B-based sparse MoE coding model with 35B total parameters, about 3B active parameters, 40 layers, 256 routed experts plus one shared expert, 8 active experts per token, 16 attention heads and 2 KV heads. The open release is text-only; the vision components are not included.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Workstation fitMac Studio M4 Max 64GBCUDA workstation classNVIDIA GB10 / DGX SparkLarge local models64GB RAM guide

Related catalogue entries

Linked mechanically by family, shared tags and nearby RAM tier; this is not a quality ranking.

Nemotron 3.5 Lightning 30B-A3B 30B (3B active, MoE)Ornith-1.5-35B-A3B 35B (3B active, MoE)Qwen 3.6 35B-A3B 35B (3B active, MoE)Qwen3-Coder-Next 80B (3B active, MoE)Qwen3.8-27B 27BMuse Glimmer 30B 29.8B multimodalAgents-A1 4B 4BKimi Linear 48B-A3B Instruct 48B (3B active, MoE)

Where to go next