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

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use unsloth/Qwen3.8-Flash-Next-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

LM Studio
Jan

vLLM

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

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Qwen3.8-Flash-Next-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/Qwen3.8-Flash-Next-GGUF",
		"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/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Ollama
How to use unsloth/Qwen3.8-Flash-Next-GGUF with Ollama:
```
ollama run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
```
Unsloth Desktop

How to use unsloth/Qwen3.8-Flash-Next-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Qwen3.8-Flash-Next-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Qwen3.8-Flash-Next-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Qwen3.8-Flash-Next-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Qwen3.8-Flash-Next-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Qwen3.8-Flash-Next-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q4_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Tested on RTX5070TI 16GB VRAM: 20 t/s

#24

by WhiteDan64 - opened about 15 hours ago

Discussion

WhiteDan64

about 15 hours ago

I was able to run Qwen3.8-Flash-Next-UD-IQ3_XXS on my RTX5071Ti (16GB) using last unsloth build: https://github.com/unslothai/llama.cpp/releases/tag/b10639-mix-f6f92fe

I reached an inference speed of about 20.44 tokens per second, not too bad for a configuration without unified memory and MTP

My presets configuration is the following:

[*]
port = 8080
host = 127.0.0.1
parallel = 1
log-verbosity = 4 
flash-attn = on
t = 18 
no-mmap = true 
metrics = true 
jinja = true 

[unsloth/qwen3.8-flash-next:UD-IQ3_XXS]
model = Qwen3.8-Flash-Next-GGUF\UD-IQ3_XXS\Qwen3.8-Flash-Next-UD-IQ3_XXS-00001-of-00003.gguf
mmproj = Qwen3.8-Flash-Next-GGUF\mmproj-BF16.gguf
no-mmproj-offload = true 
ctx-size = 262144
ctk = q4_0 
ctv = q4_0
ngl = 49 
load-mode = none
no-warmup = true
cpu-moe = true 
fit = off
spec-type = ngram-mod 
batch-size = 512
ubatch-size = 256
reasoning = auto 
temp = 0.8 
top-p = 0.90 
top-k = 20 
min-p = 0.01 
presence-penalty = 0.0 
repeat-penalty = 1.0
chat-template-kwargs = {"preserve-thinking": true, "reasoning_effort": "medium"}

Here the log

[64888] 1.33.163.298 I srv  llama_server: listening on http://127.0.0.1:64888
1.55.197.673 I srv  proxy_reques: proxying request to model unsloth/qwen3.8-flash-next:IQ3_XXS on port 64888
[64888] 1.33.177.607 I srv  update_slots: all slots are idle
1.55.227.127 I srv  proxy_reques: proxying request to model unsloth/qwen3.8-flash-next:IQ3_XXS on port 64888
[64888] 1.33.229.521 I srv  server_strea: conv_id=afebd875-5cc6-45f1-ab68-f19715b7eff8::unsloth/qwen3.8-flash-next:IQ3_XXS (empty=0)
[64888] 1.33.276.454 I srv   operator (): chat format: peg-native
[64888] 1.33.279.013 I slot get_availabl: id  0 | task -1 |  - skipping, slot is empty
[64888] 1.33.279.018 I slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
[64888] 1.33.279.019 I srv  get_availabl: updating prompt cache
[64888] 1.33.279.026 I srv          load:  - looking for better prompt, base f_keep = -1.000, f_sim = 0.000
[64888] 1.33.279.243 I srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 262144 tokens, 8589934592 est)
[64888] 1.33.279.247 I srv  get_availabl: prompt cache update took 0.23 ms
[64888] 1.33.293.296 I cmn  common_reaso: activated, budget=2147483647 tokens
[64888] 1.33.300.685 I slot launch_slot_: id  0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist
[64888] 1.33.300.715 I slot launch_slot_: id  0 | task -1 | sampler params:
[64888]         repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
[64888]         dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 64
[64888]         top_k = 20, top_p = 0.900, min_p = 0.010, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
[64888]         mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000, adaptive_target = -1.000, adaptive_decay = 0.900
[64888] 1.33.300.720 I slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
[64888] 1.33.301.378 I slot  operator (): id  0 | task 0 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 2409
[64888] 1.33.301.397 I slot  operator (): id  0 | task 0 | cached n_tokens = 0, memory_seq_rm [0, end)
[64888] 1.37.692.872 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =    512, progress = 0.21, t =   4.38 s / 117.00 tokens per second
[64888] 1.37.692.877 I slot  operator (): id  0 | task 0 | cached n_tokens = 512, memory_seq_rm [512, end)
[64888] 1.41.120.507 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   1024, progress = 0.43, t =   7.82 s / 130.96 tokens per second
[64888] 1.41.120.511 I slot  operator (): id  0 | task 0 | cached n_tokens = 1024, memory_seq_rm [1024, end)
[64888] 1.44.571.767 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   1536, progress = 0.64, t =  11.27 s / 136.29 tokens per second
[64888] 1.44.571.771 I slot  operator (): id  0 | task 0 | cached n_tokens = 1536, memory_seq_rm [1536, end)
[64888] 1.48.223.610 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   2048, progress = 0.85, t =  14.92 s / 137.25 tokens per second
[64888] 1.48.223.614 I slot  operator (): id  0 | task 0 | cached n_tokens = 2048, memory_seq_rm [2048, end)
[64888] 1.49.350.152 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   2149, progress = 0.89, t =  16.05 s / 133.90 tokens per second
[64888] 1.49.350.156 I slot  operator (): id  0 | task 0 | cached n_tokens = 2149, memory_seq_rm [2149, end)
[64888] 1.49.413.593 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 2148, pos_max = 2148, n_tokens = 2149, size = 124.876 MiB)
[64888] 1.51.001.486 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   2388, progress = 0.99, t =  17.70 s / 134.91 tokens per second
[64888] 1.51.001.490 I slot  operator (): id  0 | task 0 | cached n_tokens = 2388, memory_seq_rm [2388, end)
[64888] 1.51.049.580 I slot create_check: id  0 | task 0 | created context checkpoint 2 of 32 (pos_min = 2387, pos_max = 2387, n_tokens = 2388, size = 124.876 MiB)
[64888] 1.51.471.004 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   2405, progress = 1.00, t =  18.17 s / 132.36 tokens per second
[64888] 1.51.471.011 I slot  operator (): id  0 | task 0 | cached n_tokens = 2405, memory_seq_rm [2405, end)
[64888] 1.51.493.754 I slot init_sampler: id  0 | task 0 | init sampler, took 22.22 ms, tokens: text = 2409, total = 2409
[64888] 1.51.555.788 I slot create_check: id  0 | task 0 | created context checkpoint 3 of 32 (pos_min = 2404, pos_max = 2404, n_tokens = 2405, size = 124.876 MiB)
[64888] 1.51.916.123 I spec        begin: ngram_mod occupancy = 2353/4194304 (0.00)
[64888] 1.53.479.562 I cmn  common_reaso: deactivated (natural end)
[64888] 1.56.849.028 I slot print_timing: id  0 | task 0 | n_gen =    100, tg =  20.14 t/s, tg_3s =  20.34 t/s
[64888] 1.59.891.328 I slot print_timing: id  0 | task 0 | n_gen =    162, tg =  20.23 t/s, tg_3s =  20.38 t/s
[64888] 2.02.920.219 I slot print_timing: id  0 | task 0 | n_gen =    225, tg =  20.39 t/s, tg_3s =  20.80 t/s
[64888] 2.04.507.618 I slot print_timing: id  0 | task 0 | prompt eval time =   18631.98 ms /  2409 tokens (    7.73 ms per token,   129.29 tokens per second)
[64888] 2.04.507.623 I slot print_timing: id  0 | task 0 |        eval time =   12574.24 ms /   258 tokens (   48.93 ms per token,    20.44 tokens per second)
[64888] 2.04.507.624 I slot print_timing: id  0 | task 0 |       total time =   31206.22 ms /  2667 tokens
[64888] 2.04.507.625 I slot print_timing: id  0 | task 0 |    graphs reused =          0
[64888] 2.04.508.126 I spec common_specu: statistics        ngram-mod: #calls(b,g,a) =    1    257      0, #gen drafts =      0, #acc drafts =     0, #gen tokens =      0, #acc tokens =     0, dur(b,g,a) = 0.202, 0.590, 0.000 ms
[64888] 2.04.508.378 I slot      release: id  0 | task 0 | stop processing: n_tokens = 2666, truncated = 0

antonio20076

about 15 hours ago

Hi!
How much RAM?
Thanks!

WhiteDan64

about 14 hours ago

Total RAM 85GB (o/w 10GB Windows OS)

hgeist

about 13 hours ago

Any idea how to speed up the prefill? Currently trying to compile mtp pr and speculative prefill pr into the qwen4next pr ..

WhiteDan64

about 13 hours ago

•

edited about 12 hours ago

Hi!
How much RAM?
Thanks!

Using this preset

[*]
port = 8080
host = 127.0.0.1
parallel = 1
log-verbosity = 4 
flash-attn = on
t = 18 
metrics = true 
jinja = true 

[unsloth/qwen3.8-flash:UD-IQ3_XXS]
model = Qwen3.8-Flash-Next-GGUF\UD-IQ3_XXS\Qwen3.8-Flash-Next-UD-IQ3_XXS-00001-of-00003.gguf
mmproj = unsloth\Qwen3.8-Flash-Next-GGUF\mmproj-BF16.gguf
no-mmproj-offload = true 
ctx-size = 262144
no-context-shift = true
no-warmup = true
fit = on
spec-default = true 
batch-size = 512
ubatch-size = 256
reasoning = auto 
temp = 0.8 
top-p = 0.90 
top-k = 20 
min-p = 0.01 
presence-penalty = 0.0 
repeat-penalty = 1.0
chat-template-kwargs = {"enable_thinking": true, "preserve-thinking": true, "reasoning_effort": "low"}

Respect to my previous settings, the total RAM used is decreased to 62GB (o/w 11GB Windows OS), so probably it will be possible to run the model on 64GB (using linux you can save about 10GB)

But the prefill speed will decrease from about 130 t/s to 53 t/s, while the inference speed will increase a little from 20 t/s to 22 t/s.
So if you have enough RAM I suggest you to use my previous settings, because the prefill speed decrease too much with these settings.

khetag

about 12 hours ago

So, for IQ3_XXS, we need 62 GB of RAM and 16 GB of VRAM (78 GB total)- and is that without offloading the ngram to the SSD?

vonamakitsune

about 11 hours ago

What is the the quality though?

WhiteDan64

about 10 hours ago

•

edited about 9 hours ago

I'm still doing tests, but the preliminary results are good. Unfortunately, to increase intelligence, Qwen has greatly increased thinking time.

For example I asked to Qwen3.8-flash-next:IQ3_XXS to build a game

It was able to build a good game in 80m consuming 94k of context with reasoning effort = medium

Qwen3.8-27b:IQ3_XXS to complete the task using same prompt and effort level, took only 22m consuming 70k of context (about 26% less)

The main reason is due both the noticeable speed difference (Qwen3.8-flash-next on my PC is about 3 times slower than Qwen3.8-27b) and the increase in thinking tokens (Qwen3.8-flash-next consumes 26% more tokens than Qwen3.8-27b).

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