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

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -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

Llama new
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"

Qwen3.8-Flash-Next Unsloth Dynamic v3.1 on NVMe: a local 30-task A/B

#93

by SkyIsNotGreen - opened 6 days ago

Discussion

SkyIsNotGreen

6 days ago

•

edited 6 days ago

Hello,

I served unsloth/Qwen3.8-Flash-Next-GGUF:UD-Q2_K_XL on one workstation with:

30 GB host RAM
2 x 20 GB AMD GPUs, one of them driving the display
NVMe-backed Flash-Next weights

I did not set out to check Unsloth's internal benchmark claims. I wanted to know whether a Flash-Next quant of this class can be served usefully on a memory-constrained workstation with the sparse-expert and mmap path in Strata, and how it compares with a ternary-based Scion under the same harness.

UD-Q2_K_XL scored 30/30 (100%) and streamed 79.6 GB on the 2 x 20 GB box at about 10 effective tok/s. Scion's 60.17 GB ternary build scored 27/30 (90.0%) and ran about four times faster. The 74.28 GB Q2 bridge scored 28/30 (93.3%). This is a size and speed trade-off; I would not call it a Scion win.

Result

UD-Q2_K_XL through Strata:

Metric	Value
Tasks	30
Accuracy	30/30 (100%)
Completion tokens	11,470
Wall time	1,127.4 s
Median and p95 per task	28.6 s and 67.8 s
Effective tokens/s	10.17
ms per completion token	98.3
`finish_reason=length`	0

Same 30-task split, 8192 context, 4096 cap, 3072 thinking budget, and answer_room on:

Arm	Size	Accuracy	Tokens	Wall	Effective tok/s	ms/tok
Unsloth UD-Q2_K_XL	79.6 GB	30/30 (100%)	11,470	1,127.4 s	10.17	98.3
Scion Q2 bridge*	74.28 GB	28/30 (93.3%)	33,578	791.9 s	42.4	23.6
Scion Q6_K	60.17 GB	27/30 (90.0%)	39,922	951.4 s	42.0	23.8

* The Q2 bridge is a Strata-compatible repack of the Scion Q6_K release, not an official Scion Q2 and not ISTA's GSQ-RCO Q2_0.

The Unsloth arm had no MTP or speculative drafter (--spec 2, no --mtp). The Scion arms used spec 4 and MTP, so the speed columns are not a pure engine comparison. The accuracy columns use the same task protocol.

Accuracy vs effective throughput (completion tokens / wall time). Bubble area is file size.

Setup

Unsloth artifact: unsloth/Qwen3.8-Flash-Next-GGUF, UD-Q2_K_XL, four shards, 79.6 GB on disk. I checked all four SHA-256 hashes against the Hub.
Quant contents: mixed IQ2_XXS, IQ2_S, IQ3_XXS, and Q2_0 experts; Q8_0 and Q6_K body; IQ4_NL PLE table with 320,001,536 rows.
Pack: a local Strata tools/iq_pack.py --compat-bf16 pack. It is 1.5 GB of dense body, with 302 tensors served natively; the experts and PLE stay in the GGUF.
Engine: the local Strata fork and Flash-Next config, two AMD GPUs, PLE and n-gram table disk-backed, --mmap-experts, --prefill 4096, --spec 2, --vram-reserve-mib 6144, --vram-reserve-later-mib 1024, --pool-workers 15.
Harness: the same 30-task split and checkers used for the Scion rows above.
Token budget: max_tokens=4096, thinking_budget_tokens=3072, answer_room on.
temperature=0.1, one sample per task, no thinking-preservation, raw OpenAI-compatible chat completions.
The task prompts ask for direct answers and final-answer markers. This is a local A/B benchmark, not a re-implementation of Unsloth's methodology.

How it ran

The Flash-Next PLE and n-gram table stays on disk instead of being duplicated into RAM.
Expert banks are read from the GGUF shards. The pack does not need a full copy of the model.
Strata's expert cache keeps hot expert banks in VRAM. The 79.6 GB model ran on 40 GB of VRAM plus the 30 GB host, with the rest read from NVMe.
The local pack conversion is not bit-identical to stock llama.cpp or Strata builds that read the file natively. 291 small non-native tensors were converted to the engine's form with --compat-bf16 and nearest-even rounding: 96 exact, 195 rounded, maximum absolute error 0.0154. The conversions are logged in compat-bf16.json and conversions.json.

Streaming

Average decode was 12.09 tok/s and average prompt was 44.09 tok/s on the 2 x 20 GB box. On the same box, the 60.17 GB Scion Q6_K ran at about 42 effective tok/s, roughly four times the throughput.
The 30-task run took 1,127 s end to end with 11,470 completion tokens. Decode waited on disk, which is the cost of streaming the 79.6 GB file.

Accuracy by benchmark, every task's wall time, and total completion tokens.

Caveats

I ran one server arm at a time. Both GPUs are used by the layer split, and the setup is not tuned for maximum throughput.
This is not a comparison against Unsloth's published long-thinking 256k-context benchmarks.
This is not one of Unsloth's recommended xhigh reasoning runs. It is a one-sample, low-temperature matrix meant to expose code, reasoning, and QA failures.
The Unsloth arm ran without MTP or speculation; the Scion arms used spec 4 and MTP. Speed is not apples-to-apples.
The Scion and Unsloth arms used different VRAM reserves (3072 and 6144) and expert-cache placements. The 30-task suite is small and cache-sensitive.
The pack build rounds small tensors. Do not quote it as a stock or official production benchmark.
One 30-task run is not frontier capability evidence.
The download is large enough that normal machines will not run this casually. The question here is whether this tool lets a Flash-Next-scale model run usefully on constrained hardware.

Reproduction

Harness: hivebench/experiments/ab_qwen27_flashnext.py and hivebench/experiments/model_matrix.py in sky-is-green/hive
Test split: hivebench/experiments/cascade/tasks-bench.json
Raw data: hivebench/experiments/model-matrix/ab-unsloth-flashnext-ud-q2k-xl-4096-tb3072.json, ab-scion-q6k-v0210-4096-tb3072.json, ab-scion-q2-v0210-4096-tb3072.json
Unsloth config and stack: hivebench/stacks/strata/configs/unsloth-flashnext-ud-q2_k_xl.json, hivebench/stacks/unsloth-flashnext-strata.json
Pack conversion logs: compat-bf16.json and conversions.json beside the pack. The SHA-256 hashes for all four shards were verified against the Hub.

Why I'm posting this

I'm posting this because it's a concrete example of Flash-Next-scale inference on hardware that does not hold the full model in RAM. The PLE table and expert pages stay on disk, hot experts go to vRAM, and the same task harness shows whether the local implementation produces useful outputs. On this 30-task matrix the Unsloth quant wins on accuracy. Scion's value is that it reaches 90% at 60 GB and roughly four times the throughput of the streamed 79.6 GB model. Both are constrained-hardware results, not contradictions of the model's high-end serving behavior.

License and attribution

Upstream weights: Qwen3.8-Flash-Next by Qwen, and the Unsloth Dynamic v3.1 GGUF quant from unsloth.
Local tooling: sky-is-green/Strata, a Strata and Prism llama.cpp derivative, plus the community harness and prompt set in sky-is-green/hive.
Scion model and training: sky-is-green/scion and SkyIsNotGreen/Scion-FlashNext-176B-A6B.
This is an independent community result. No vendor endorsed it.

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