license: other
license_name: qwen-community-license-1.0
license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE
base_model:

• Qwen/Qwen3.8-Flash-Next
  pipeline_tag: text-generation
  library_name: gguf
  tags:
• gguf
• gufo
• strix-halo
• qwen
• uncensored
• abliterated
• MoE
• requantized
• amd-rocm

Qwen3.8-Flash-Next Uncensored — Q4_K_XL-Gufo

gufo-compatible requantization of the community "Abliterated" (uncensored)
fine-tune of Qwen3.8-Flash-Next — prepared for the
gufo inference engine on AMD
Strix Halo (gfx1151).

Created 2026-10-23 from the existing 3-shard GGUF
Qwen3.8-Flash-Next-Uncensored-Q4_K_S — no new fine-tuning run, purely a
layout/metadata correction that makes the same weights readable by gufo.

Lineage

1. Qwen/Qwen3.8-Flash-Next — base model (architecture qwen4exp: 48 blocks,
   512 experts, size label 512x56B, 262 144 native context)
2. Community fine-tune Uncensored / Abliterated (AI red-team) —
   orcarouter/Qwen3.8-Flash-Next-Uncensored
3. GGUF export Q4_K_S —
   orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
   (3 shards, ~105 GiB, imatrix quantization)
4. This requantization for gufo (see "What was changed")

What was changed (and why)

The gufo engine validates a per-tensor format whitelist and the
index-ring geometry while creating the model. The original Q4_K_S export
failed both and was rejected by gufo. This copy passes:

Quality note: the upgraded tensors contain no new information — they are
the same weights as the Q4_K_S export, just stored loss-free where gufo demands
high-precision input. Net effect vs. the original GGUF: equal to slightly better
(critical tensors are no longer hard-quantized to q4_k).

Files

Qwen3.8-Flash-Next-Uncensored-Q4_K_XL-Gufo-00001-of-00003.gguf  (39.7 GB)
Qwen3.8-Flash-Next-Uncensored-Q4_K_XL-Gufo-00002-of-00003.gguf  (47.7 GB)
Qwen3.8-Flash-Next-Uncensored-Q4_K_XL-Gufo-00003-of-00003.gguf  (23.7 GB)

Usage (gufo)

podman run --rm --name gufo --device /dev/dri --device /dev/kfd \
  --group-add keep-groups --security-opt seccomp=unconfined --ipc=host \
  --security-opt label=disable --userns=keep-id -p 8028:8028 \
  -v <MODEL_DIR>:/models/target:ro \
  docker.io/kyuz0/amd-strix-halo-toolboxes:rocm-10.0-gufo \
  gufo serve --host 0.0.0.0 --port 8028 --sessions 2 llm \
  --model /models/target/Qwen3.8-Flash-Next-Uncensored-Q4_K_XL-Gufo-00001-of-00003.gguf \
  --context 262144 --max-tokens 32768 --served-model-name Qwen3.8-Flash-Next \
  --think on --reasoning-effort high

How this model runs in the AI Toolbox ecosystem (headless autostart via
file mode, discrete model swap, LiteLLM) is documented in
docs/AI-TOOLBOX-SETUP.md — including what the
cockpit itself can and cannot do with gufo.

Full provenance: the exact per-tensor dtype tables used for the requantization
ship in docs/ — gufo-requant-minimal-rules.txt (the 716 exact-name rules
actually applied), gufo-requant-tensortypes-exact.txt (the complete 1224-entry
reference table read from the UD-Q4_K_XL shards) and gufo-requant-tensortypes.txt
(the condensed pattern list).

License — please read (documented license discrepancy)

The licensing of this model is not as clean as the embedded metadata suggests:

How we handle it. Legally decisive is the license of the source repositories:
the weights derive from the Qwen base model, and a "mere" conversion or
requantization cannot grant more rights than the upstream license allows. Therefore
this copy is redistributed under the terms of the Qwen Community License 1.0 —
its full text ships in LICENSE. The apache-2.0 claim embedded by
the GGUF exporter is documented here (and kept untouched in the metadata), but it
cannot override the upstream license text for the weights. The lineage is preserved
in the GGUF metadata (general.base_model.*).

Good news: the Qwen Community License 1.0 is essentially MIT with model-specific
additions — it explicitly permits use, modification, fine-tuning, distribution and
sale of the weights. Conditions: (1) keep the copyright/permission notice (the
prominent-name clause only applies above 100 M monthly active users or USD 20 M
monthly revenue); (2) a separate Qwen license is required only for commercial
"Model-as-a-Service" or "AI Work Assistant" businesses. Private and local use —
like this one on Strix Halo — only carries the notice requirement.

Disclaimer: documented as found, not legal advice. A friendly question to OrcaRouter
or Qwen clears the last ambiguity.

Acknowledgements

• @kyuz0 — ai-toolbox-cockpit
  and the Strix Halo toolbox images that run gufo here. Many thanks!
• gufo-org — for the engine
• the Qwen team — for the base model
• Unsloth — whose validated UD-Q4_K_XL layout served as the reference recipe
• @orcarouter — for the uncensored/abliterated fine-tune and the original
  GGUF export on which this requantization is based