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
```

Strix Halo teaser: 88 GB Qwen3.8-Flash-Next (512 experts) at 45–50 tok/s on a Radeon 8060S — what we've been cooking in our llama.cpp fork

#70

by Fbrbiasi - opened Sep 29

Discussion

Fbrbiasi

Sep 29

•

edited Oct 2

We've been quietly grinding on the Ryzen AI Max+ 395 (Radeon 8060S, gfx1151, 128 GB unified) for a few weeks. Here's a peek before the full write-up.

The model: Qwen3.8-Flash-Next. 177B MoE with ~3B active, 512 experts, top-10, hybrid Gated DeltaNet + sparse attention. We run unsloth's UD-IQ4_XS GGUF (87 GiB).

What it does on one APU, measured end to end with real ~8.5K and ~34K-token prompts:

45–50 tok/s decode with MTP speculative decoding (draft 3, 66–70% acceptance)
27–28 tok/s without speculation
~~1,200 tok/s prefill, even on 34K-token prompts (~~1,100 with the MTP head loaded)
For scale: upstream llama.cpp on the same machine, with the same GGUF and prompts, does ~220–320 tok/s prefill and 14–19 tok/s decode (its build has no MTP for this architecture)
Running daily as the backend of our local agent setup, with conversations of 50K+ tokens

What's under the hood (a few highlights):

Next on the bench: the larger UD-Q4_K_XL quant (111 GB), which cuts unsloth's measured KLD from 0.084 to 0.047. We want to see what that quality costs in speed and memory on a 128 GB machine.

Also coming: our sub-2-bit ternary formats (TQ1_1 / TQ2_1). A 27B model in 5.5–6.7 GB, running on a 6 GB GPU from 2016 or on an iGPU with no dedicated card.

All numbers are from a single PX13 laptop with strix halo
on AC power: engine-reported, median of warm runs, the same prompts for every engine, temperature 0. Full methodology, the op-by-op ROCm vs Vulkan breakdown, and quality numbers (KLD vs the reference quant) are coming in the proper post.

It will be available on 10/09/26 https://github.com/ftoleedo

What would you want to see benchmarked on Strix Halo? Drop your model and quant below.

vstiff

Sep 29

Since the model's context is really compact per token, so I'd like to see a bit higher memory utilization, like 100Gb. Currently running Unsloth UD-Q4_K_XL, fits pretty tight, but slower that your numbers here.

Fbrbiasi

Sep 30

•

This comment has been hidden (marked as Resolved)

Fbrbiasi

Oct 1

my last tests

Fbrbiasi changed discussion title from Strix Halo teaser: 88 GB Qwen3.8-Flash-Next (512 experts) at 42–44 tok/s on a Radeon 8060S — what we've been cooking in our llama.cpp fork to Strix Halo teaser: 88 GB Qwen3.8-Flash-Next (512 experts) at 45–50 tok/s on a Radeon 8060S — what we've been cooking in our llama.cpp fork Oct 2

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