Instructions to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
# Run inference directly in the terminal:
llama cli -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
# Run inference directly in the terminal:
llama cli -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
# Run inference directly in the terminal:
./llama-cli -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

Use Docker

docker model run hf.co/AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

LM Studio
Jan

vLLM

How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AnonimousA/Qwen3.8-Flash-Next-REAP-320-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": "AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

Ollama
How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with Ollama:
```
ollama run hf.co/AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
```
Unsloth Desktop

How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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": "AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with Docker Model Runner:
```
docker model run hf.co/AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL
```

Lemonade

How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

Run and chat with the model

lemonade run user.Qwen3.8-Flash-Next-REAP-320-GGUF-UD-Q2_K_XL

List all available models

lemonade list

Hermes Agent

How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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 AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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 "AnonimousA/Qwen3.8-Flash-Next-REAP-320-GGUF:UD-Q2_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"

Having problem running with MTP enabled

by hkhan777 - opened about 1 month ago

Discussion

hkhan777

about 1 month ago

Hi,

first of all, thanks for the great modal! I am getting around 24 tok/s on 7900xtx with llama-cpp build with ROCm 7.1x.
I have tried running it with unsloth 8b MTP head as suggested in description but I get error of parameter mismatch

/build/bin/llama-cli
-m ~/.cache/huggingface/hub/models--AnonimousA--Qwen3.8-Flash-Next-REAP-320-GGUF/snapshots/5fb294f5d45410d31a3dca15277d87a19b659630/Q2/Qwen3.8-Flash-Next-UD-Q2_K_XL-reap320-00001-of-00002.gguf -ngl 52 -fa on --parallel 1 -c 100000 -ctv q4_0 -ctk q4_0 --temperature 1.0 --top_p 0.95 --top_k 20 --min_p 0.0 --presence_penalty 0.0 -t 16 --n-cpu-moe 28 -md ~/Downloads/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf --spec-type draft-mtp --spec-draft-n-max 2 --no-mmap

Loading model... |0.00.759.084 E llama_model_load: error loading model: check_tensor_dims: tensor 'token_embd.weight' not found
0.00.759.095 E llama_model_load_from_file_impl: failed to load model
|0.12.013.181 E llama_model_load: error loading model: check_tensor_dims: tensor 'token_embd.weight' not found
0.12.013.185 E llama_model_load_from_file_impl: failed to load model
0.12.013.186 E common_speculative_init_result: failed to load draft model, '/Downloads/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf'
0.12.013.195 E srv load_model: failed to load draft model, '/Downloads/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf'
0.12.013.947 E srv llama_server: exiting due to model loading error
/llama_server exited with code 1
Error: the server exited before becoming ready

It would be great if you could guide me. Thanks in Advance!

AnonimousA

Owner about 1 month ago

Thanks for the report. 24 tok/s on a 7900 XTX with 28 expert layers on CPU is a solid number for this model.

The error is expected on mainline llama.cpp, and the file is fine. The shared MTP heads deliberately omit token_embd.weight (and the output projection): they borrow both from the main model at load time, which is what saves ~1.3 GB. Mainline has the --spec-type draft-mtp flag, but it does not have the Qwen3.8-Flash-Next MTP graph nor the "borrow from the target" loader, so it tries to open the draft as a standalone model and stops at the first missing tensor. That is exactly the message you got.

Fix: use a build that actually has the MTP support. Three options, all listed in Unsloth's MTP/README.md:

Unsloth's fork, release tag b10715-mix-86bd2d3 or newer: https://github.com/unslothai/llama.cpp/releases. The prebuilt binaries there are CUDA/CPU only, so for ROCm take the source tarball of that tag and build it the same way you built mainline (-DGGML_HIP=ON). This is what we run here (a b10798-mix build on Windows/CUDA).
Upstream PR https://github.com/ggml-org/llama.cpp/pull/28243 (same code on its way to mainline, still marked draft as of today).
Unsloth fork PR #144.

Once on one of those builds keep exactly the flags you already have (-md … --spec-type draft-mtp --spec-draft-n-max 2). At startup the shared head prints one error line about borrowing the embeddings. That line is normal, it works.

Don't try the non-shared mtp-Qwen3.8-Flash-Next-Q8_0.gguf on mainline as a workaround: it carries its own embeddings so it gets past this error, but mainline still has no MTP graph for this architecture, so it fails later.

Two things we measured that may help on a 24 GB card:

The head costs ~3 GB of VRAM no matter what --n-cpu-moe is, so expect to push 2–3 more expert layers to CPU to fit. The shared-Q4_K_M head saves another ~840 MiB with the same acceptance (we measured 0.70 vs 0.72 for Q8_0), worth it at 24 GB.
Acceptance is 0.7–0.8 with n-max 2 and net decode gain on our side was about +30%. Another user reported --spec-draft-n-max 3 working well for him, so it is worth a quick A/B once it loads.

Let me know what you get once it is running.

hkhan777

about 1 month ago

Thanks for the detailed guidance, I was able to use Unsloth's shared MPT head with their llama.cpp fork as you suggested and got upto 30 tok/s at 8K context.

However, I tried a frontend task, a casual Kanban prompt from https://github.com/ChapouX/benchmarking/blob/main/prompts/kanban.txt got very bizarre UI, everything worked but the UI was like web 1.0.

First I ran same command with thinking on and got similar result, it looked more functional than this but all vertical again. Maybe you can try this as well?

hkhan777 changed discussion status to closed about 1 month ago

hkhan777 changed discussion status to open 30 days ago

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