license: other
license_name: qwen-community-1.0
license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE
base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored
base_model_relation: quantized
tags:
- exl3
- exllamav3
- qwen3.8
- flash-next
- abliterated
- uncensored
- mtp
- moe
- 4-bit
- strix-halo
- crucible
language: - en
- fr
- zh
pipeline_tag: image-text-to-text
Qwen3.8-Flash-Next-Uncensored: EXL3 4.04 bpw (crucible build)
EXL3 4-bit quant of
orcarouter/Qwen3.8-Flash-Next-Uncensored,
the refusal-removed build of Qwen3.8-Flash-Next, re-spliced for the
Crucible inference engine on AMD
Strix Halo (gfx1151). Built 2026-10-24. Derived from
Lygodactylus/Qwen3.8-Flash-Next-Uncensored-exl3-4bpw.
Crucible is the only engine this exact file set loads under. Anything else should
use the parent pack.
| Base model | orcarouter/Qwen3.8-Flash-Next-Uncensored (179.55B-param MoE, text + vision + MTP head) |
| Format | EXL3 trellis, 4.0416 bpw measured trunk, 6-bit lm_head, BF16 token embedding |
| Size on disk | ~100 GiB: 9 weight shards (67.5 GB) + 39 GB n-gram (PLE) table + 0.56 GB vision tower |
| Weights in memory | ~67 GB (n-gram table paged, not resident) |
| Cold boot to /health | about 110 s on a 128 GB Strix Halo APU |
What is different from the parent 4bpw pack
Crucible's loader enforces three things stock ExLlamaV3 does not, and this pack
closes all of them in the weights.
- MTP head. The parent keeps its MTP block at 8 bpw, which the engine
rejects. The MTP tensors from the 4.05 bpw
base pack are
spliced into the shard stream at 4, 5 and 6 bpw, so the pack is
self-contained. Load it without--mtp-sidecar. - Fused MoE kernel contract. The engine takes routed experts at 3 or 4 bpw
and shared experts at routed+2. Layer-0 routed experts were re-encoded K5 to
K4, and all 140 non-compliant shared experts K5 to K6. - Vision tower. The parent's raw BF16 tower (one fused-qkv scan with a
4304-wide MLP) is replaced by the 6-bp tower the parent also ships
(vision_k6.safetensors).
The 1,676 re-encodes are per-tensor quantize_exl3 (uncalibrated meta-H) from
the pristine BF16 source with ExLlamaV3 v1.4.4, ref-checked against BF16 before
splicing: relative Frobenius error 0.017, correlation 0.99985. The parent pack's
stock encodings measured 3.65 percent on the same probe, so the re-encodes land
closer to source than the parent.
Measured on the released files
| Trellis widths | 4 bpw x 75,265 tensors, 5 bpw x 104, 6 bpw x 382 |
| MTP block | mixed 4, 5 and 6 bpw |
| Vision tower | 6 bpw |
| Trunk bpw | 4.0416 |
config.json declares bits: 4.04 and mtp_bits: 4. The MTP field is a
nominal floor, because the MTP bytes carry three widths per the table; the
trunk declaration matches measurement at two decimals.
Checks run before release
- Extended structural check on the final pack: OK, zero problems (tensor
duplicates, MoE widths, vision geometry). - Fresh engine boot:
/healthok, a sanity completion answered correctly, the
vision tower loads (0.35 GB). - The abliteration survived the splice. Four refusal prompts that the unmodified
base lane refuses four out of four times are all released by this pack, and a
control prompt plus an image description come back identical on both lanes.
No model-quality benchmark numbers on this card. Any throughput figure quoted
elsewhere was taken on other hardware and other quantizer versions, so it would
mislead here.
Running it
Point Crucible on Strix Halo at
the pack directory.
- The MTP head is in the weights. No sidecar argument.
- Give each pack its own conversation-state directory. A saved state is a KV
cache built against one pack's weights, and resuming it on a different pack
means decoding with the wrong trunk. - Under llama-swap, set the lane's
useModelNametocrucible, the engine's
own model id. Any other name gets a 404.
llama.cpp, Ollama, vLLM and stock ExLlamaV3/TabbyAPI cannot load this pack
as-is; use the parent pack for those.
Uncensored
Safety alignment has been substantially removed by abliteration, and the splice
did not weaken it (the checks above). There are no meaningful built-in
guardrails. Add your own moderation layer before any deployment.
Credits
- orcarouter/Qwen3.8-Flash-Next-Uncensored:
the abliteration and the BF16 source - Lygodactylus:
the parent quant this pack is derived from - turboderp:
ExLlamaV3 and the base pack whose MTP sidecar feeds the splice - The splice and the 1,676 re-encodes were built on-device on Strix Halo, 2026-10-24
License
Qwen Community License 1.0, inherited from the base model, same terms as the
parent pack.