KAT-Coder-V2.5-Dev

KAT-Coder-V2.5-Dev

An open-weight coding model from Kwaipilot, a mixture-of-experts with 35B total and 3B active parameters, tuned for agentic coding and fewer tool-call glitches. Works with transformers, vLLM and SGLang.

Made by Kwaipilot on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: KAT-Coder-V2.5-Dev is the open-weight version of Kwaipilot's KAT-Coder V2.5 coding model. It is a mixture-of-experts model with 35B parameters in total and 3B active, post-trained with supervised learning and reinforcement learning.

The authors say it reaches state-of-the-art results for agentic coding among models of similar size. They also report that reinforcement learning cut odd behaviors, such as malformed tool labels falling from 9.34% to 0.28%.

Who it suits: Developers and teams who want an open coding model for agent workflows on their own servers, and researchers who compare coding agents at this size.

Industry | Who uses it | How they use it | Example
Software | Developer-tool teams | Power a coding agent with an open model | An in-house assistant that edits repositories
Research | Agent researchers | Compare coding models of similar size | Running SWE-bench style tests
Platforms | Inference teams | Serve an open model with a standard engine | Hosting it through vLLM or SGLang

You can skip it if you want a general chat model or a small model for a laptop, since this one is aimed at coding agents.

Where to use it: The card says the weights are in Transformers format and work with several engines, in English and Chinese.

• Hugging Face Transformers
• vLLM
• SGLang
• KTransformers

Use case | Who | What they get back
Agentic coding | Developers | Code edits and tool calls from an agent loop
Repository fixes | Engineering teams | Patches for issues, as in SWE-bench style tasks
Bilingual coding help | Teams in English and Chinese | Answers in both languages

How to set it up: You need Python, a GPU server and one of the supported engines. The card gives only the download step in the visible part, and the weights are about 69.3 GB in BF16.

• Python with pip
• Roughly 69.3 GB for BF16 weights, or about 17.3 GB at 4-bit as an estimate
• One of Transformers, vLLM, SGLang or KTransformers

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Load the model in your chosen engine.

pip install -U huggingface_hub
hf download Kwaipilot/KAT-Coder-V2.5-Dev

How to use it: Serve or load the model in your engine and drive it from a coding agent. The card shows no run command, and it notes that the open release ships only the language-model weights.

Example: You connect the model to an agent framework and ask it to fix a failing test in a repository. The model reads the files, calls its tools and returns a patch.

For newcomers and experts: If you are new, this is a model for coding agents, not for casual chat. Download the files and load them in a standard engine before you try a small repository task.

If you know the field, only 3B of its 35B parameters are active per token, which keeps serving costs closer to a small model. The card credits SFT and RL training and reports lower rates of abnormal tool labels and repetition.

Post-train with SFT and RL
Load in Transformers, vLLM or SGLang
Agent sends a coding task
Model calls tools and edits code
Patch returned to the agent

Good to know: The card compares the model with several open models. On SWE-bench Verified it scores 69.40, the highest of the five models shown here.

Title: SWE-bench Verified score
KAT-Coder-V2.5-Dev: 69.40 points
Qwen3.5-27B: 68.60 points
Qwen3.6-35BA3B: 64.40 points
Gemma4-31B: 60.60 points
Qwen3.5-35BA3B: 58.60 points

• License: Apache 2.0
• Cost to run: none named, but a 35B model needs a capable GPU server
• Data it reaches: the repositories and files you give the agent, and the model files you download
• Not for: tasks beyond the language model, since the open release ships only the language-model weights

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 34.7B
• Weights: about 69.3 GB in BF16; about 17.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 16,106
• Likes: 687
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 73
• Fine-tunes: 12
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, code, agent, agentic-coding, coding, text-generation, conversational

Links

• Model: https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev
• Files: https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Kwaipilot/KAT-Coder-V2.5-Dev

Then follow the usage in its model card. Nothing in this kit runs by itself.