halogen-qwen3.8-flash-next

halogen-qwen3.8-flash-next

Holds the model weights for halogen-flash-server, an engine that runs Qwen3.8-Flash-Next on AMD Strix Halo machines, for owners of that hardware who want the model running locally.

Made by peonist-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: The files here feed halogen-flash-server, a purpose-built engine for one job: serving Qwen3.8-Flash-Next on the AMD Strix Halo chip, known as gfx1151. The files use halogen's own .hgn format, so transformers, vLLM and llama.cpp cannot load them. You mount them into a halogen-flash-server container so the engine can read them.

Who it suits: It suits people who own an AMD Strix Halo machine with a large amount of memory and want to run Qwen3.8-Flash-Next on it themselves.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Self-hosting | Home lab owners | Run a large model on one AMD box | Serve the model from a Strix Halo machine
Software | Developers | Use a local model server in their workflow | Point an app at the local server's port
AI infrastructure | Engineers | Test an inference engine tuned for one chip | Compare the checkpoint options on their own machine

If you do not have a Strix Halo machine, or you want a standard model format, you can skip it.

Where to use it:

• The halogen-flash-server container, run with podman or Docker
• An AMD Strix Halo host, the card notes a 128 GB one
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Local text generation | Developers | A model server running on your own machine
Image input | Users who send screenshots | Image reading, once the optional vision file is added
Lower memory option | Owners with tight memory | A smaller checkpoint that holds about 8 GiB less

How to set it up: You need a Strix Halo machine with plenty of memory and about 110 GiB of disk space. At run time about 62 GiB sits in memory, and the lookup table is read through memory mapping on top of that.

• Python with pip, for the hf command
• podman or Docker
• An AMD Strix Halo machine, ideally one that is free for this job

1. Download the repository into a models folder.
2. Start the server container with the AMD GPU devices passed in and the models folder mounted read-only.
3. On Docker instead of Podman, swap the keep-groups option for the video and render groups.
4. Optional: set HALOGEN_VISION_TOWER=1 to turn on image input, or HALOGEN_CHECKPOINT to pick the smaller checkpoint.

hf download peonist-ai/halogen-qwen3.8-flash-next --local-dir ~/halogen-models

How to use it: Once the container is up, the server listens on port 8731 and you send it requests from your apps. The card shows the container command in full, and it also tells you which files are needed for each path.

Title: Files in the repository
Tool | What it does
qwen38-flash-next-v2.hgn | The default checkpoint, 62.1 GiB
qwen38-flash-next-ngram.hgn | Its lookup table, which you must keep, 47.7 GiB
qwen38-flash-next-ht43.hgn | Optional smaller checkpoint, 53.7 GiB
qwen38-flash-next-vision.hgn | Optional file for reading images, 0.84 GiB
qwen38-flash-next-mtp.hgn | Draft head for running a GGUF copy, 1.42 GiB

Example: You download the repository, start the container, and send a text prompt to port 8731. With the vision file beside the checkpoint and the vision option on, you can also send a screenshot for the model to read.

For newcomers and experts: This is a specialist setup, so start by checking that you have the hardware. A first small step is downloading the files and reading the card's file list before starting the container.

The ht43 option compresses the expert layers harder, saving roughly 8 GiB of memory, and the price is somewhat slower prompt processing and a small drop in decoding speed. Since version 0.7.0 the server can also open a llama.cpp GGUF directly, and the draft-head file supplies the head that a GGUF lacks. Older w4b files from version 0.14 still load.

Download the .hgn files
Mount them into the container
halogen-flash-server loads the model
Requests arrive on port 8731

Good to know: Give the server a machine to itself. With default settings it claims most of a 128 GB host's memory, and other big programs running there can freeze for minutes.

Title: File sizes in the repository
v2 checkpoint: 62.1 GiB
Lookup table: 47.7 GiB
ht43 checkpoint: 53.7 GiB
w4b checkpoint: 115.55 GiB
Draft head: 1.42 GiB
Vision file: 0.84 GiB

• License: Apache 2.0.
• Cost to run: none named, apart from the AMD hardware.
• Data it reaches: the files come from Hugging Face, and the model then runs on your own machine.
• Not for: transformers, vLLM or llama.cpp, or hardware other than AMD Strix Halo.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: halogen
• Downloads in the last 30 days: 0
• Likes: 198
• License: apache-2.0
• Based on: Qwen/Qwen3.8-Flash-Next
• Fine-tunes: 1
• Tags: halogen, halogen-flash, qwen, mixture-of-experts, long-context, quantized, 4-bit, strix-halo, gfx1151, rocm, radeon, ryzen-ai

Links

• Model: https://huggingface.co/peonist-ai/halogen-qwen3.8-flash-next
• Files: https://huggingface.co/peonist-ai/halogen-qwen3.8-flash-next/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download peonist-ai/halogen-qwen3.8-flash-next

Then follow the usage in its model card. Nothing in this kit runs by itself.