Instructions to use Qwen/Qwen3.8-Flash-Next with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen3.8-Flash-Next with Transformers:

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

pipe = pipeline("image-text-to-text", model="Qwen/Qwen3.8-Flash-Next")
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)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Qwen/Qwen3.8-Flash-Next")
model = AutoModelForMultimodalLM.from_pretrained("Qwen/Qwen3.8-Flash-Next", 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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Qwen/Qwen3.8-Flash-Next with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen3.8-Flash-Next"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen3.8-Flash-Next",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen3.8-Flash-Next

SGLang

How to use Qwen/Qwen3.8-Flash-Next 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 "Qwen/Qwen3.8-Flash-Next" \
    --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": "Qwen/Qwen3.8-Flash-Next",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "Qwen/Qwen3.8-Flash-Next" \
        --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": "Qwen/Qwen3.8-Flash-Next",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen3.8-Flash-Next with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen3.8-Flash-Next
```

Model cannot SPLIT TENSOR and will not work in consumer multi gpu

#62

by Dr-Ojinsky - opened Sep 19

Discussion

Dr-Ojinsky

Sep 19

0.00.407.274 I print_info: file size = 157.54 GiB (7.65 BPW)
0.00.415.334 E llama_model_load: error loading model: LLAMA_SPLIT_MODE_TENSOR not implemented for architecture 'qwen4exp'
0.00.415.345 E llama_model_load_from_file_impl: failed to load model
0.00.415.351 E cmn common_init_: failed to load model 'c:\AI_Models\Qwen3.8-Flash-Next-UD-Q6_K_XL-00001-of-00006.gguf'
0.00.415.355 E srv load_model: failed to load model, 'c:\AI_Models\Qwen3.8-Flash-Next-UD-Q6_K_XL-00001-of-00006.gguf'
0.00.415.361 I srv operator(): operator(): cleaning up before exit...
0.00.416.547 E srv llama_server: exiting due to model loading error

Nerdsking

Sep 19

•

edited Sep 19

Yes, this is happening to my also. I got two rtx 3080 with 20GB VRAM each. If I use "-fit on" and llama uses both gpus, the final speed is worse than if I were using just one GPU. My CPU is a ryzen 3700x with 128GB VRAM 3600mhz

llama version used: b11040
layer split doesn't work.

pdrolet

29 days ago

I have tried an unsloth version Q2_K_XL on my MacBook Pro M5 max (128gb). It uses about 80gb of ram. And it is VERY good (an understatement!). And it quite fast too. Very impressive. I have used this model for my meeting summaries and it is way better than any local model I have tried so far. I have not tried it for coding, but I will.

donaldJJ

22 days ago

The error only concerns -sm tensor. The current server README lists it as tensor: split weights and KV across GPUs (parallelized, EXPERIMENTAL), and it hasn't been implemented for qwen4exp. -sm layer is the default and does load. On consumer cards, tensor split wouldn't be the win anyway: most of this model has to sit in system RAM, so the offload split matters much more than how the GPUs share the work.

@Nerdsking some estimates for 2×20 GB + 128 GB DDR4-3600, using unsloth's file sizes, 64K context at f16 KV and ~1.5 GiB of buffers:

UD-IQ4_XS (93.7 GB): one GPU needs -ncmoe 39 → ~18 GiB VRAM, ~45 GiB RAM. Two GPUs need only -ncmoe 21 → ~38.5 GiB VRAM, ~25 GiB RAM.
UD-Q4_K_XL (111 GB): one GPU -ncmoe 41, two GPUs -ncmoe 27 (~41 GiB RAM).
On top of that, the ~26.8 GiB n-gram table stays on the CPU side. It's read lazily from the mmapped file.

Decode is bound by what's read from RAM per token (10 of 512 experts active). Dual-channel DDR4-3600 is ~57.6 GB/s. For IQ4_XS that gives roughly 12–20 tok/s on one card vs 16–27 tok/s on two, so the second card should help, not hurt.

If two GPUs came out slower, I'd check two things:

Is --fit (on by default) actually moving fewer experts to the CPU when both GPUs are visible? If it keeps the same -ncmoe, you pay the pipeline hop and gain nothing.
With -ncmoe N set by hand, layers 0..N-1 are light on the GPU. Under layer split, GPU0 gets mostly light layers unless you adjust -ts.

Setting -ncmoe explicitly plus -ts and comparing llama-bench numbers against one GPU should show which one it is.

For the Q6_K_XL (157.5 GiB): 40 GiB VRAM + 128 GB RAM ≈ 159 GiB total, so it won't realistically fit next to the OS.

I got these per-file numbers from a calculator I built for -ncmoe: https://modelvram.com/moe-offload-calculator/ (I built this). They're bandwidth estimates, not measurements.

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