Instructions to use shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed with Transformers:

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

pipe = pipeline("text-generation", model="shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed")
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, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed")
model = AutoModelForImageTextToText.from_pretrained("shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed")
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 shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed

SGLang

How to use shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed 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 "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed" \
    --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": "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed",
		"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 "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed" \
        --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": "shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed with Docker Model Runner:
```
docker model run hf.co/shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed
```

Qwen3.6-35B-A3B PARO full4096-e5 — packed format

This is the packed ParoQuant export for Qwen/Qwen3.6-35B-A3B, using the full4096-e5 calibration run.

The packed artifact was produced from the legacy/original export with:

python3 scripts/strip_paro_safetensors.py \
  --input-dir /models/qwen36-quant/Qwen3.6-35B-A3B-PARO-full4096-e5 \
  --output-dir /models/qwen36-quant/Qwen3.6-35B-A3B-PARO-full4096-e5-packed \
  --mode packed

Packed changes:

Removed every duplicate fp16 .weight fallback tensor where the same module has .qweight
Removed tensors: 250
Removed tensor bytes: 2,810,183,680
model.safetensors: 20,474,495,512 bytes
Actual packed BPW: 4.6799 using a 35B denominator
Verified duplicate fallback count after stripping: 0

The legacy/original-format export is available separately at:

shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5

Quality and size comparison

Canonical tx4/quality3 evaluation compares each candidate directly against the original BF16 HF model on the same scored token positions.

Model	Format	Size GiB ↓	BPW ↓	PPL ↓	ΔNLL ↓	KL nats ↓	Top-1 % ↑
Original BF16 HF	HF safetensors	66.966	16.435	6.5590	+0.000000	0.000000	100.000
PARO full4096-e5 packed (this repo)	packed safetensors	19.068	4.680	6.6216	+0.009506	0.034684	92.000
PARO full4096-e5 unpacked/original-format	legacy safetensors	21.686	5.322	6.6216	+0.009506	0.034684	92.000
GGUF UD-Q4_K_S	GGUF	19.458	4.776	6.5783	+0.003842	0.012800	94.999
GGUF UD-Q4_K_M	GGUF	20.614	5.059	6.5643	+0.001718	0.010849	95.354

Column calculations:

Size GiB: active model weight artifact bytes divided by 2^30; for HF/PARO this sums .safetensors weight files only, and for GGUF this is the .gguf file size.
BPW: active model weight artifact bytes × 8 / 35,000,000,000; the fixed 35B denominator is used only to make formats directly comparable.
PPL: exp(mean NLL) over the canonical scored target positions only. Evaluation source: tx4/quality3 validation set, ctx=2048, stride=1023, 127 windows, 129,921 scored tokens.
ΔNLL: candidate mean NLL minus original BF16 HF mean NLL; positive values mean worse true-token likelihood than the original.
KL nats: mean KL(P_original_BF16_HF || P_candidate) over the full next-token distribution; lower is better.
Top-1 %: percentage of scored positions where the candidate and original BF16 HF choose the same argmax next token; higher is better.
Packed and unpacked PARO quality metrics are identical because the packed artifact removes duplicate fp16 fallback tensors only; the quantized tensors are unchanged.
GGUF rows were evaluated with llama.cpp using BF16 HF reference log-probabilities streamed in llama.cpp KLD format.

Notes

This artifact requires a packed-aware ParoQuant-compatible loader/runtime; legacy loaders that expect duplicate fp16 fallback .weight tensors will not load this format.

See strip_paro_safetensors_report.json for the exact stripping report.

Downloads last month: 74

Safetensors

Model size

6B params

Tensor type

I32

F16

I16

Model tree for shisa-ai/Qwen3.6-35B-A3B-PARO-full4096-e5-packed

Base model

Qwen/Qwen3.6-35B-A3B

Quantized

(420)

this model