G1LL1/Qwen3.8-Flash-Next-Uncensored-orca-halogen

Qwen3.8-Flash-Next-Uncensored — Halogen-compatible checkpoint

This repository provides a native Halogen checkpoint converted from the IQ4_XS quantization of OrcaRouter’s Qwen3.8-Flash-Next-Uncensored-GGUF.

The checkpoint is intended for halogen-flash-server on AMD Strix Halo (gfx1151).

About the conversion

The file was produced using Halogen’s standalone convert mode. According to Halogen’s documentation, this lossless repack preserves the source GGUF’s quantized tensor values without requantizing them.

The resulting .hgn contains the repacked language-model weights, the n-gram lookup table, and Halogen’s compatible MTP draft head from peonist-ai/halogen-qwen3.8-flash-next.

• Source quantization: IQ4_XS
• Distributed format: native Halogen .hgn
• Checkpoint: qwen3.8-flash-next-uncensored.hgn
• Tokenizer: downloaded separately below
• Vision: optional, using a separate Halogen vision tower

No additional fine-tuning or abliteration is performed by this conversion. It retains the source model’s behavior and limitations.

The original GGUF shards are unnecessary when serving this complete checkpoint. HALOGEN_GGUF_CACHE and a separate MTP-head file are also unnecessary. Halogen’s quality overlay for its own W4B checkpoint does not apply to this converted model.

The .hgn file is intended for Halogen and cannot be loaded as a GGUF in llama.cpp or Ollama.

Requirements

• Native Linux with an appropriate AMD driver configuration.
• AMD Strix Halo hardware with GPU architecture gfx1151.
• Sufficient unified memory for the weights, runtime allocations, and context cache. Halogen’s reference system has 128 GB.
• Podman with access to /dev/kfd and /dev/dri.
• A recent Hugging Face CLI (hf).
• Enough local storage for the checkpoint and optional vision tower.

Consult the Halogen documentation for host configuration and memory requirements.

Download and start the text server

If this repository is private, first run hf auth login with an account that has access.

The following downloads the checkpoint and the required tokenizer, then starts the server:

mkdir -p "$HOME/models/orca-halogen"

hf download G1LL1/Qwen3.8-27B-Uncensored-orca-halogen \
  qwen3.8-flash-next-uncensored.hgn \
  --local-dir "$HOME/models/orca-halogen"

hf download peonist-ai/halogen-qwen3.8-flash-next \
  --include "tokenizer/*" \
  --local-dir "$HOME/models/orca-halogen"

podman run --rm \
  -p 127.0.0.1:8731:8731 \
  --device /dev/kfd \
  --device /dev/dri \
  --group-add keep-groups \
  --ipc=host \
  --ulimit memlock=-1:-1 \
  -v "$HOME/models/orca-halogen:/models:ro" \
  -e HALOGEN_CHECKPOINT=/models/qwen3.8-flash-next-uncensored.hgn \
  -e HALOGEN_TOKENIZER=/models/tokenizer \
  -e HALOGEN_MODEL_ID=qwen3.8-flash-uncensored \
  ghcr.io/peonist-ai/halogen-flash-server:latest

If the checkpoint and tokenizer are already present, run only the podman run command.

The OpenAI-compatible API is available at http://localhost:8731/v1. These examples bind the API to localhost.

Download and enable vision

After completing the downloads above, stop the text-only server before starting this alternative configuration.

Download the Halogen vision tower, approximately 898 MB, and enable it explicitly:

hf download peonist-ai/halogen-qwen3.8-flash-next \
  qwen38-flash-next-vision.hgn \
  --local-dir "$HOME/models/orca-halogen"

podman run --rm \
  -p 127.0.0.1:8731:8731 \
  --device /dev/kfd \
  --device /dev/dri \
  --group-add keep-groups \
  --ipc=host \
  --ulimit memlock=-1:-1 \
  -v "$HOME/models/orca-halogen:/models:ro" \
  -e HALOGEN_CHECKPOINT=/models/qwen3.8-flash-next-uncensored.hgn \
  -e HALOGEN_TOKENIZER=/models/tokenizer \
  -e HALOGEN_MODEL_ID=qwen3.8-flash-uncensored \
  -e HALOGEN_VISION_TOWER=/models/qwen38-flash-next-vision.hgn \
  ghcr.io/peonist-ai/halogen-flash-server:latest

HALOGEN_VISION_TOWER=1 is also supported when the vision file is beside the checkpoint under its original filename. This setting discovers an existing file; it does not download it.

Halogen accepts image inputs through /v1/chat/completions and /v1/responses. Send images as base64 data URLs or supported base64 content. Remote HTTP(S) image URLs are not fetched by this runtime.

Verify the server

From another terminal:

curl --fail --silent --show-error http://localhost:8731/health
curl --fail --silent --show-error http://localhost:8731/v1/models

For the standalone converted checkpoint, /health should report checkpoint_format: "hgn". With the vision tower loaded, it should also report vision.enabled: true.

The configured API model ID is qwen3.8-flash-uncensored.

License and attribution

The original Qwen3.8-Flash-Next is published under the Qwen Community License 1.0. Conversion, quantization, and abliteration do not remove applicable upstream license obligations.

Although some downstream model repositories label their artifacts Apache-2.0, those metadata labels should not be interpreted as waiving the original Qwen license.

The Qwen license permits use, modification, and redistribution subject to its conditions, including preservation of its copyright and permission notices. It also contains specific commercial-use provisions:

• Licensees or affiliates conducting a Model as a Service or AI Work Assistant business generally require a separate Qwen license for commercial use, subject to the stated internal-use exception.
• Certain large commercial products or services must prominently display the model name.

Refer to the complete upstream license for the definitions, thresholds, and conditions.

The Halogen inference engine is licensed separately under the Peonist EULA. The model license does not license the runtime.

Credits:

• Qwen: original model.
• OrcaRouter: uncensored model and source GGUF quantization.
• Peonist: Halogen runtime, conversion tooling, compatible MTP head, and optional vision tower.
• G1LL1: checkpoint conversion.

This is an independent community conversion. Compatibility references do not imply endorsement by the upstream authors.