Instructions to use Kwaipilot/KAT-Coder-V2.5-Dev with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Kwaipilot/KAT-Coder-V2.5-Dev with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Kwaipilot/KAT-Coder-V2.5-Dev")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Kwaipilot/KAT-Coder-V2.5-Dev")
model = AutoModelForMultimodalLM.from_pretrained("Kwaipilot/KAT-Coder-V2.5-Dev", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Kwaipilot/KAT-Coder-V2.5-Dev with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Kwaipilot/KAT-Coder-V2.5-Dev"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Kwaipilot/KAT-Coder-V2.5-Dev",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Kwaipilot/KAT-Coder-V2.5-Dev

SGLang

How to use Kwaipilot/KAT-Coder-V2.5-Dev with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Kwaipilot/KAT-Coder-V2.5-Dev" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Kwaipilot/KAT-Coder-V2.5-Dev",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Kwaipilot/KAT-Coder-V2.5-Dev" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Kwaipilot/KAT-Coder-V2.5-Dev",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Kwaipilot/KAT-Coder-V2.5-Dev with Docker Model Runner:
```
docker model run hf.co/Kwaipilot/KAT-Coder-V2.5-Dev
```

Kwaipilot wins!

#27

by Tip-Tutaki - opened 2 days ago

Discussion

Tip-Tutaki

2 days ago

Blind comparison against kwaipilot kat coder v2.5 dev, qwen 3.6 35b a3b, glm 4.7 flash, qwen 2.5 coder, ornith 1.0 9b, granite 4.1 8b, qwen 3.5 4b showed kat coder 95/100 for oneshotting a n8n workflow from a 25 word prompt. second was qwen 3.6 35b a3b at 90/100, granite. GLM 5.2 was the judge

This is an absolute goldmine of information, and it completely contextualizes why we saw the results we did!

First of all, the biggest plot twist here is that KAT-Coder-V2.5-Dev is literally built on top of Qwen3.6-35B-A3B. So when you said "you proved that kwaipilots finetune is actually better than qwen 3.6," you were 100% spot on—we were literally watching the base model go head-to-head with its own fine-tuned offspring, and the fine-tune absolutely cleaned its clock.

The technical report is fascinating, but a few things specifically stand out that perfectly explain why Model 1 performed so flawlessly on your n8n prompt:

1. The "Excessive Parallel Tool Calls" Penalty
The report mentions that base Qwen3.6 had a nasty habit of spamming up to 70 parallel tool calls in a single turn, causing context bloat and training collapse. Kwaipilot specifically applied RL penalties to fix this.
Why this matters for your prompt: The n8n qBittorrent workflow requires strict, sequential logic (Login -> Get Torrents -> Move Files -> Delete). A model trying to call all these APIs in parallel would fail miserably. KAT-Coder's RL training specifically taught it to follow clean, sequential execution paths without hallucinating shortcuts.

2. Hierarchical Rewards based on Harness Execution
Instead of just giving the model a 0 or 1 based on whether the final code compiled, they rewarded "meaningful progress in unsuccessful trajectories."
Why this matters for your prompt: This trains the model to think like a developer. It’s why KAT-Coder didn't just write the API calls, but also realized, "Wait, if I tell qBittorrent to delete the torrent with deleteFiles=true, it will wipe the files I just moved." It anticipates the execution environment's state.

3. Token-in-Token-out (TITO) Consistency
This is a brilliant detail. A lot of models fail at generating complex n8n JSON because the chat template or tokenizer mangles the whitespace and escaping in the code blocks during inference. By ensuring strict TITO consistency, KAT-Coder guarantees that the massive JSON block it generated for your workflow was perfectly formatted and escaped, making it genuinely ready to copy-paste.

4. The Benchmark Flex
Seeing it beat Qwen3.6 by ~5 points on SWE-bench Verified (69.40 vs 64.40) and absolutely dominate on Scicode (44.20 vs 37.53) is massive for a 3B active parameter MoE.

It’s also hilarious to read their callout about Qwen3.5-35BA3B hallucinating tools (MultiEdit) that don't exist. That is exactly the behavior we saw in the lower-scoring models in your test (hallucinating native n8n nodes or qBittorrent webhooks that don't exist).

Kwaipilot basically took an already smart base model, put it through a brutal RL bootcamp to cure its bad habits (repetition, tool spamming, hallucinations), and turned it into an absolute monster of a coding agent. Thanks for sharing the model card—this is seriously impressive engineering!

MartinPatterson

1 day ago

How about against ornith 1.0 35b? Woudln't that be more relevant comp?

Tip-Tutaki

1 day ago

I want to start by saying I was excited when I saw ornith 1.0 35b come out and read through its model card as a potential replacement for qwen 3.6 35b a3b. Unfortunately it did not get there. It was due to the ornith models failing me that I went seeking other fine tuned of qwen 3.6 35b a3b and I came across kwaipilot Kat coder dev.

From my own experiences and other independent testing that I have been seeing, the model card numbers for ornith 1.0 35b weren't relevant to the model they released.

On another note I will be conducting my own testing with kwaipilot v2.5 dev in the q6k and q8 quantizations over the weekend. I plan to run a full suite of benchmarks/evals across it in unsloth studio chat, openclaw, Hermes agent, opencode, qwencode and vscode.

tooltd

1 day ago

Based on my personal experience, ranking is as follows: Qwen3.8 27B 🤤 > Fable Fusion 711 > 27B Base > kwaipilot v2.5 dev > Ornith > 35B base
Agents-A1 hasn't been tested yet.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment