Instructions to use sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP with Transformers:

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

pipe = pipeline("text-generation", model="sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP")
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("sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP")
model = AutoModelForMultimodalLM.from_pretrained("sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP")
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 sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP

SGLang

How to use sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP 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 "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP" \
    --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": "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP",
		"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 "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP" \
        --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": "sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP with Docker Model Runner:
```
docker model run hf.co/sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP
```

Broken chat_template tool calling and thinking

by livepeer-ren - opened Apr 28

Discussion

livepeer-ren

Apr 28

I was having a lot of problems with silent ends of my coding agents. It seems to do with chat_template as per https://www.reddit.com/r/Qwen_AI/comments/1stt081/fixed_jinja_chat_templates_for_qwen_35_and_36/
https://github.com/abysslover/qwen36_tool_calling_failure/tree/main
So I think it is worth to include this fixes into the repo.

livepeer-ren

Apr 28

This comment has been hidden (marked as Low Quality)

sakamakismile

Owner Apr 29

Thanks for the report and for pointing to the Qwen3.6 chat template/tool-calling issue.

I reviewed the shipped chat_template.jinja and confirmed that these Qwen3.6-27B-family NVFP4 repos were still using the upstream Qwen3.6 template, so they could inherit the reported coding-agent/tool-calling and thinking-toggle problems.

I have prepared and rolled out a conservative patched chat_template.jinja for the affected Qwen3.6-27B-family releases. The patch keeps Qwen's existing XML-like tool-call format and preserves the original vision/text rendering, while adding fixes for:

leading developer messages by folding them into the system block
<|think_on|> / <|think_off|> handling without exposing those control tokens to the model
enable_thinking=false behavior
historical assistant.tool_calls serialization for both mapping and JSON-string arguments
including the tool/function name in tool responses when available

I also tested the template rendering locally for normal chat, VLM/image content, tool calls with mapping arguments, tool calls with JSON-string arguments, and thinking-off cases before applying it.

Thanks again for catching this.

— Tonoken3 / LNA-LAB

livepeer-ren

Apr 29

Thanks for the report and for pointing to the Qwen3.6 chat template/tool-calling issue.

I reviewed the shipped chat_template.jinja and confirmed that these Qwen3.6-27B-family NVFP4 repos were still using the upstream Qwen3.6 template, so they could inherit the reported coding-agent/tool-calling and thinking-toggle problems.

I have prepared and rolled out a conservative patched chat_template.jinja for the affected Qwen3.6-27B-family releases. The patch keeps Qwen's existing XML-like tool-call format and preserves the original vision/text rendering, while adding fixes for:

leading developer messages by folding them into the system block

<|think_on|> / <|think_off|> handling without exposing those control tokens to the model

enable_thinking=false behavior

historical assistant.tool_calls serialization for both mapping and JSON-string arguments

including the tool/function name in tool responses when available

I also tested the template rendering locally for normal chat, VLM/image content, tool calls with mapping arguments, tool calls with JSON-string arguments, and thinking-off cases before applying it.

Thanks again for catching this.

— Tonoken3 / LNA-LAB

Thanks for help!! Does it also include fix for this? https://huggingface.co/Qwen/Qwen3.6-27B/discussions/16

sakamakismile

Owner Apr 29

Quick follow-up: I also checked the Qwen3.6-35B-A3B-family NVFP4 releases.

The official Qwen/Qwen3.6-35B-A3B chat_template.jinja is byte-identical to the Qwen3.6-27B template, so the same tool-calling / thinking-toggle issues can be inherited there as well. I have now rolled out the same conservative patched template to the affected 35B-A3B-family repos too.

This keeps the upstream vision/text rendering and Qwen XML-like tool-call format, while adding the same fixes for leading developer messages, <|think_on|> / <|think_off|>, enable_thinking=false, historical tool-call argument serialization, and function names in tool responses.

— Tonoken3 / LNA-LAB

livepeer-ren

Apr 29

•

edited Apr 29

does it also apply this? https://huggingface.co/Qwen/Qwen3.6-27B/discussions/20 It is known bug that qwen halucinate block into from time to time and this should be addressed as well.

livepeer-ren

Apr 29

This comment has been hidden (marked as Low Quality)

livepeer-ren

Apr 29

This comment has been hidden (marked as Low Quality)

livepeer-ren changed discussion status to closed Apr 29

livepeer-ren

Apr 29

Quick follow-up: I also checked the Qwen3.6-35B-A3B-family NVFP4 releases.

The official Qwen/Qwen3.6-35B-A3B chat_template.jinja is byte-identical to the Qwen3.6-27B template, so the same tool-calling / thinking-toggle issues can be inherited there as well. I have now rolled out the same conservative patched template to the affected 35B-A3B-family repos too.

This keeps the upstream vision/text rendering and Qwen XML-like tool-call format, while adding the same fixes for leading developer messages, <|think_on|> / <|think_off|>, enable_thinking=false, historical tool-call argument serialization, and function names in tool responses.

— Tonoken3 / LNA-LAB

i am testing your fixed chat template and so far i see MUCH more robust chats, no silent drops yet!! Thanks

livepeer-ren

Apr 29

livepeer-ren changed discussion status to open Apr 29

livepeer-ren

Apr 29

•

edited Apr 29

I had issues with lack of prompt caching... 0%. I changed over to this chat template from https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4 and all is good now.
Edit: idk now it is fine prompt caching is at 80% so probably skill issue on my end.
Edit; redhat template is stopping silently as wel..
Now using this fixed template https://github.com/allanchan339/vLLM-Qwen3.5-27B/blob/main/qwen3.5-enhanced.jinja
and --tool-call-parser qwen3_xml . Going for 2 hrs without silent stop on opencode.

livepeer-ren

Apr 30

https://www.reddit.com/r/LocalLLM/comments/1sv6cqk/follow_up_tested_tool_calling_fixes_for_qwen/ THIS IS WORKING WITH LATEST VLLM NIGHTLY DOCKER IMAGE

sakamakismile

Owner Apr 30

Hi @livepeer-ren ,

Thank you for the screenshot and for the Reddit reference. I read through the follow-up post as well.

The distinction around preserve_thinking is especially useful. It sounds like there are two related but different serving paths that should not be mixed accidentally:

The current Qwen3.6 / vLLM nightly path using reasoning_parser=qwen3, tool_call_parser=qwen3_coder, and reasoning separation.
The qwen3.5-enhanced.jinja path, where preserve_thinking=false is mandatory, and qwen3_coder is preferred for Qwen3.6 because it can catch tool calls even when the model leaves a thinking block unclosed.

That is a very helpful clarification. I’ll avoid documenting this as a single universal recipe until I compare both paths more carefully.

The driver / NCCL notes from the Reddit post are also important, especially the warning that NVIDIA Studio Driver 595.79 deadlocks can look like tool-calling failures. I’ll keep that in mind when testing and documenting the recommended setup.

For now, I’ll treat the latest vLLM nightly Docker image plus qwen3_coder as the most promising direction, and I’ll test whether the model repos should keep the conservative template, adopt the enhanced template, or document both variants depending on preserve_thinking.

Thanks again. This is exactly the kind of real agentic workflow report that helps make the model card recommendations practical.

Best,
Tonoken3 / LNA-LAB

livepeer-ren

Apr 30

•

edited Apr 30

I just run another test. With qwen_xml and this worked 180k context agentic task in opencode without any hiccups. chat_template.jinja from https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/discussions/2 with this fix to also work with openclaw. Works well finally! latest image: vllm/vllm-openai:nightly is mandatory! also with --preserve thinking true

Bellesteck

May 3

Quick follow-up: I also checked the Qwen3.6-35B-A3B-family NVFP4 releases.

The official Qwen/Qwen3.6-35B-A3B chat_template.jinja is byte-identical to the Qwen3.6-27B template, so the same tool-calling / thinking-toggle issues can be inherited there as well. I have now rolled out the same conservative patched template to the affected 35B-A3B-family repos too.

This keeps the upstream vision/text rendering and Qwen XML-like tool-call format, while adding the same fixes for leading developer messages, <|think_on|> / <|think_off|>, enable_thinking=false, historical tool-call argument serialization, and function names in tool responses.

— Tonoken3 / LNA-LAB

I can confirm that the base model also has this issue. It's kinda rare - I see it maybe 2 to 6 times per 1000 requests.

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