Huihui Qwen3.8-Flash-Next abliterated (HGN v2)
huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated converted directly from its BF16 safetensors into the HGN v2 weight format used by halogen-flash-server (AMD Strix Halo, gfx1151).
The source is an uncensored version of Qwen/Qwen3.8-Flash-Next made with abliteration. This is not the stock model: for the stock model use peonist-ai's own halogen-qwen3.8-flash-next.
[!WARNING] Uncensored model. Its safety filtering has been significantly reduced, so it may produce sensitive, controversial or inappropriate content and is not suitable for all audiences. You are solely responsible for how you use it and for complying with applicable laws. The source authors recommend research or controlled use, monitoring and reviewing outputs, and avoiding direct use in public-facing commercial applications.
These weights are in .hgn format and will not load in transformers, vLLM or llama.cpp. They are meant to be mounted into the halogen-flash-server container.
Status (read this first)
This is a community conversion. It has not been qualified by the authors of any upstream project.
| Check | Result |
|---|---|
Conversion (bf16-to-hgn convert, profile v2, HF BF16 input) |
Completed and published: 1,658 source tensors to 2,349 output tensors, manifest included |
bf16-to-hgn verify (source hashes, layout, decoded values) |
Completed: provenance attested, no tensor above 5% error (see below) |
| Loads and generates in halogen-flash-server | Not yet recorded |
| Speed and quality benchmarks | Not yet recorded |
| CUDA output compared byte-for-byte with the CPU backend | Not done. The converter documents CUDA as qualified on sm_89; this run used sm_86 (RTX 3090) |
If you test it, please report results on the discussion tab.
Files
| File | Size | Purpose |
|---|---|---|
model.hgn |
66,687,678,432 B (62.1 GiB) | Checkpoint, identity qwen3.8-flash-next-v2. Includes the MTP draft head |
qwen38-flash-next-ngram.hgn |
51,200,246,084 B (47.7 GiB) | The 51B-parameter n-gram lookup table (FP8). Required; keep beside model.hgn |
qwen38-flash-next-vision.hgn |
897,916,416 B (0.84 GiB) | Vision tower (BF16). Optional; without it the engine is text-only |
config.json |
Model configuration, unchanged from the source | |
preprocessor_config.json, video_preprocessor_config.json |
Processor configuration, copied from the source | |
tokenizer/ |
Flat tokenizer directory (also contains config.json) |
|
conversion-manifest.json |
Converter manifest: policy, spec version, source provenance and per-file details | |
LICENSE |
Qwen Community License 1.0, which must accompany copies of the weights |
Total about 111 GiB. The model needs about 62 GiB resident plus the memory-mapped n-gram table.
How it was made
Qwen/Qwen3.8-Flash-Next base model
-> huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated abliteration, BF16 safetensors
-> this repo HGN v2 via bf16-to-hgn
- Converter: Quat3rnion/bf16-to-hgn with the HF input path, profile
v2, policybf16-hgn-v2-rtn-nf4-signed-rms-trellis-v1,--ht-beam 128, seed 0. Format per hgn-spec 1.1.0. - Everything comes from the source snapshot: the converter's HF path takes the MTP draft head, the vision tower, the n-gram table and the tokenizer assets from the supplied snapshot, so no stock sidecar was substituted.
- Backend: CUDA (
--ht-backend cuda --i4r-backend cuda --q4cp-backend cuda) on an RTX 3090. The conversion passes (231.2G values) took 23 minutes 43 seconds.
Reproduce:
bf16-to-hgn convert --source /path/to/Huihui-Qwen3.8-Flash-Next-abliterated \
--output Huihui-Qwen3.8-Flash-Next-abliterated-HGN \
--profile v2 --threads 8 --ht-beam 128 \
--ht-backend cuda --i4r-backend cuda --q4cp-backend cuda
bf16-to-hgn verify --source /path/to/Huihui-Qwen3.8-Flash-Next-abliterated \
--output Huihui-Qwen3.8-Flash-Next-abliterated-HGN --decode-backend cuda
Conversion error (from bf16-to-hgn verify)
The source is BF16, so these figures are measured against the original weights. They are the relative error energy of the output's decoded values, which is quantization error and not a quality score.
| Tensor family | Error |
|---|---|
| Routed experts (I4R) | 0.939% in every layer (0.9387–0.9394%) |
| LM head / token embeddings | 0.75% / 0.67% |
| Attention and linear-attention projections | 0.62–0.69% |
| Shared experts | 0.61–0.62% |
in_proj_a, in_proj_b, shared-expert gate |
0.8–1.1% |
| Hyper-connection mixers (Q6G64) | about 0.07% |
| N-gram table (FP8) | 0.07% |
| MTP head | routed experts 0.68%, shared-expert gate 1.03%; its other tensors are below 0.07% |
| 52 of 100 tensor families | exactly 0 (the BF16/F32 pass-through tensors, such as norms, router, SSM state and the vision tower) |
The verification run reports provenance_attested: true, 1,658 source tensors, 2,349 output tensors and 118,785,459,868 payload bytes. No tensor exceeds 5% error.
Precision: what this is and is not
HGN v2 is a 4-bit-class format: routed experts use I4R, dense matrices and the LM head use HT4 trellis encoding, hyper-connection mixers use Q6G64, and the router, norms and vision tower stay BF16. Converting from BF16 means every weight is quantized once, from the source values. The converter uses a fixed deterministic policy with no calibration data, and it makes no claim of equivalence to the source or to the stock checkpoint. No perplexity or benchmark comparison has been run for this conversion.
Speculative decoding
The checkpoint includes the MTP draft head from the source snapshot. Per the halogen documentation, speculative decoding is byte-identical to serial greedy decoding, so a poorly matched head costs speed, not correctness. It can be disabled per request.
Use with halogen-flash-server
Requires an AMD Strix Halo machine (gfx1151) with enough unified memory for roughly 62 GiB of weights plus the memory-mapped table, and a halogen-flash-server release of 0.15 or later. Mount the downloaded directory read-only, as for any halogen checkpoint:
hf download johnlockejrr/Huihui-Qwen3.8-Flash-Next-abliterated-HGN --local-dir ~/halogen-models
podman run --rm -p 8731:8731 \
--device /dev/kfd --device /dev/dri --group-add keep-groups \
--ipc=host --ulimit memlock=-1:-1 \
-e HALOGEN_CHECKPOINT=/models/model.hgn \
-e HALOGEN_VISION_TOWER=1 \
-v ~/halogen-models:/models:ro \
ghcr.io/peonist-ai/halogen-flash-server:<version>
Notes:
- Replace
<version>with a current server release.HALOGEN_VISION_TOWER=1is optional and only needed for images. HALOGEN_CHECKPOINTpoints atmodel.hgnby name because the checkpoint is not namedqwen38-flash-next-v2.hgn. This checkpoint has not been loaded in the engine yet. If the engine does not find the n-gram file beside the checkpoint, see the server documentation.- On Docker, replace
--group-add keep-groupswith--group-add video --group-add render. - The server exposes an OpenAI-compatible API on port 8731.
Credits
- Base model: Qwen/Qwen3.8-Flash-Next
- Abliterated weights: huihui-ai, made with remove-refusals-with-transformers
- HGN format specification: jtsylve/hgn-spec
- Converter: Quat3rnion/bf16-to-hgn
- Engine and the HGN format it uses: peonist-ai (halogen-flash-server)
- HGN conversion: johnlockejrr
This repository is an independent community conversion and is not affiliated with or endorsed by Qwen, huihui-ai, Peonist or any other project above.
License
The weights are a derivative of Qwen3.8-Flash-Next and are provided under the Qwen Community License 1.0 (Copyright (c) 2026 Qwen); the full text is in LICENSE and the copyright and permission notice must be kept with all copies. In summary, as the license states: commercial products or services with more than 100,000,000 monthly active users or US$ 20,000,000 monthly revenue must prominently display the model name in the user interface, and a licensee that conducts a Model as a Service or AI Work Assistant business needs a separate license from Qwen before using the Software or its derivatives for any commercial purpose. This summary is not legal advice; read LICENSE and verify the terms yourself before commercial use.
The halogen-flash engine is separate software distributed under its own terms. "halogen" is a trademark of Peonist, LLC; see TRADEMARKS.md in the server repository for permitted use.