Instructions to use iromu/Qwen3-0.6B-tools-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use iromu/Qwen3-0.6B-tools-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="iromu/Qwen3-0.6B-tools-NVFP4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("iromu/Qwen3-0.6B-tools-NVFP4")
model = AutoModelForCausalLM.from_pretrained("iromu/Qwen3-0.6B-tools-NVFP4", device_map="auto")
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 iromu/Qwen3-0.6B-tools-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "iromu/Qwen3-0.6B-tools-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "iromu/Qwen3-0.6B-tools-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/iromu/Qwen3-0.6B-tools-NVFP4

SGLang

How to use iromu/Qwen3-0.6B-tools-NVFP4 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 "iromu/Qwen3-0.6B-tools-NVFP4" \
    --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": "iromu/Qwen3-0.6B-tools-NVFP4",
		"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 "iromu/Qwen3-0.6B-tools-NVFP4" \
        --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": "iromu/Qwen3-0.6B-tools-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use iromu/Qwen3-0.6B-tools-NVFP4 with Docker Model Runner:
```
docker model run hf.co/iromu/Qwen3-0.6B-tools-NVFP4
```

Qwen3-0.6B Tools NVFP4

NVFP4 quantization of the Qwen3-0.6B model fine-tuned for tool calling and agent-style interactions.

Base model

This model was fine-tuned from:

Qwen/Qwen3-0.6B

Quantization

Quantized to NVFP4 (FP4 e2m1 weights + FP8 e4m3 block scales, group size 16) with NVIDIA ModelOpt 0.46.0 (NVFP4_DEFAULT_CFG, W4A4, max scaling), calibrated on 32 samples of the sft_tools train split.

NVFP4 is native to NVIDIA Blackwell (GB10 / DGX Spark).

File	Dtype
`model.safetensors`	FP4 (packed U8) weights, FP8 e4m3 scales

Training

Training was performed using NVIDIA NeMo AutoModel with LoRA/PEFT.

LoRA configuration

LoRA rank (dim): 32
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: *.proj

Training configuration

Sequence length: 4096
Micro batch size: 2
Gradient accumulation: 32
Effective batch size: 64
Steps: 336
Learning rate: 5e-5
Weight decay: 0.01
Optimizer: AdamW
Loss: masked cross entropy
Reasoning content masking: disabled

Dataset

The model was trained on the sft_tools portion of:

r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation

The dataset contains tool/agent interaction examples.

Intended use

This model is intended for:

tool calling
function calling
agent workflows
structured tool interactions
local AI assistants
lightweight local inference

It is not intended to be a general replacement for larger Qwen models.

Serving with TensorRT-LLM

Example:

trtllm-serve serve iromu/Qwen3-0.6B-tools-NVFP4 --port 8000

The checkpoint is loaded directly — no conversion step is needed.

Validation matrix

Tool-calling validation on the sft_tools validation split (greedy decoding, 384 max new tokens). Throughput is single-stream greedy decode, not serving throughput.

Pretrained base (Qwen/Qwen3-0.6B): 1.1% exact-args match (3/274). Fine-tuned (BF16): 66.0% exact-args match (33/50) (+64.9pp vs base).

NVFP4: 0/2 (0.0%) exact, - tok/s — 0% of BF16.

Model	Quant	n	Tool call emitted	Names match	Exact args match	Δ exact vs BASE	tok/s
Qwen3-0.6B-tools	BASE (Qwen/Qwen3-0.6B)	274	51/274 (18.6%)	4/274 (1.5%)	3/274 (1.1%)	—	35.5
Qwen3-0.6B-tools	BF16	50	49/50 (98.0%)	43/50 (86.0%)	33/50 (66.0%)	+64.9pp	33.3
Qwen3-0.6B-tools	NVFP4	2	2/2 (100.0%)	0/2 (0.0%)	0/2 (0.0%)	-1.1pp	-