Instructions to use saricles/Qwen3-Coder-Next-NVFP4-GB10 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use saricles/Qwen3-Coder-Next-NVFP4-GB10 with Transformers:

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

pipe = pipeline("text-generation", model="saricles/Qwen3-Coder-Next-NVFP4-GB10")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("saricles/Qwen3-Coder-Next-NVFP4-GB10")
model = AutoModelForCausalLM.from_pretrained("saricles/Qwen3-Coder-Next-NVFP4-GB10")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use saricles/Qwen3-Coder-Next-NVFP4-GB10 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "saricles/Qwen3-Coder-Next-NVFP4-GB10"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "saricles/Qwen3-Coder-Next-NVFP4-GB10",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/saricles/Qwen3-Coder-Next-NVFP4-GB10

SGLang

How to use saricles/Qwen3-Coder-Next-NVFP4-GB10 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 "saricles/Qwen3-Coder-Next-NVFP4-GB10" \
    --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": "saricles/Qwen3-Coder-Next-NVFP4-GB10",
		"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 "saricles/Qwen3-Coder-Next-NVFP4-GB10" \
        --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": "saricles/Qwen3-Coder-Next-NVFP4-GB10",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use saricles/Qwen3-Coder-Next-NVFP4-GB10 with Docker Model Runner:
```
docker model run hf.co/saricles/Qwen3-Coder-Next-NVFP4-GB10
```

NVFP4-GB10 production feedback + reproducible-bench setup — recipe?

by nooneknows1 - opened 13 days ago

Discussion

nooneknows1

13 days ago

Hi Michael,

Quick context: I'm running your Qwen3-Coder-Next-NVFP4-GB10 in production on a DGX Spark via LocalAI + vLLM 0.23.0 with FlashInfer-Cutlass kernels. Real workload is a Hermes-CLI coding agent doing tool-call sessions, hitting ~62 tok/s steady-state. Solid build — thanks for the work.

I'm currently building a small reproducible benchmark suite (TTFT, throughput, HumanEval-pass-rate, tool-call compliance) that compares NVFP4 builds against each other under realistic streaming-with-tools workloads. The plan is to feed the results back to anyone whose build I include. As context for the benchmark itself: I just landed a vLLM streaming PR for progressive emission with active tool parsers (mudler/LocalAI#10351, with E2E numbers against your build).

Would you be open to sharing your llm-compressor recipe (oneshot/Quant config + recipe yaml/py)? I'd like to make sure I'm reproducing your build exactly as the baseline before testing variant calibrations (code-focused datasets, larger sample counts, ignore-list tweaks). Happy to keep it private if you'd prefer not to publish it publicly yet.

In return I can share the benchmark harness once it's stable and feed quality-numbers back to you per build — useful signal if you want to publish better quality cards on the models.

If you'd rather not share the recipe — totally fine, I'll reconstruct from the model-card hints. Mostly wanted to ask first since you've got the ground-truth setup.

Cheers,
Philipp

nooneknows1

13 days ago

Closed — switching to a private DM instead. Apologies for the noise.

nooneknows1 changed discussion status to closed 13 days ago

nooneknows1

13 days ago

Re-opening — apologies for the noise.

nooneknows1 changed discussion status to open 13 days ago

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