Qwen3.8-Flash-Next-Uncensored - EXL3 5.05 bpw

EXL3 quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored, an abliterated (refusal-removed) build of Qwen/Qwen3.8-Flash-Next. Built with ExLlamaV3 1.4.6.

Same recipe as the 4.05 bpw build, one step up.


Size on disk	124.5 GB (116 GiB)
of which n-gram table	39.0 GB
of which model weights	85.4 GB
Language model	5.05 bpw
`lm_head`	6 bpw
MTP layers	8 bpw
n-gram (PLE) tables	6 bpw
Vision tower	unquantized (16-bit)

The n-gram table is identical in size to the 4.05 bpw build's (same 6 bpw setting); all the extra size is in the model weights.

Sibling repos

Variant	Repo
3 bpw	Qwen3.8-Flash-Next-Uncensored-exl3-3bpw
4.05 bpw	Qwen3.8-Flash-Next-Uncensored-exl3-4bpw
5.05 bpw	this one

Config

model:
  model_name: Qwen3.8-Flash-Next-Uncensored-exl3-5bpw
  cache_size: 16384
  ngram_ram: true
  tensor_parallel: false  # true needs ExLlamaV3 1.5.1+

draft_model:
  draft_mode: mtp

tensor_parallel needs ExLlamaV3 1.5.1 or later for Qwen4ExpForConditionalGeneration (QSA layers are not split). On earlier versions it fails with NotImplementedError: Tensor-parallel is not currently implemented for Qwen4ExpForConditionalGeneration and you must use autosplit. Not tested on my side.

ngram_ram: true loads the n-gram table into system RAM instead of streaming it from disk per forward.

See the 4.05 bpw card for notes on cache quantization (needs ExLlamaV3 1.4.8 or later, and buys little on this model).

Conversion

python convert.py \
  -i Qwen3.8-Flash-Next-Uncensored \
  -o Qwen3.8-Flash-Next-Uncensored-exl3-5bpw \
  -w /tmp/exl3-work \
  -b 5.05 -hb 6 -ngb 6 -mb 8 -vb 16 \
  -d 0,1,2,3 -v

Default calibration corpus, 250 rows x 2048 columns.

Measurements

None. The model weights alone (85.4 GB, n-gram table excluded) exceed the 80 GB of VRAM on my test box (4x RTX 4000 Ada), so this build was converted there but not run. Speed figures for the same recipe at 4.05 bpw are on the 4.05 bpw card.

Quality

Not benchmarked at 5.05 bpw. Abliteration figures are orcarouter's, measured on the BF16 source, not on this quant.

Safety

Safety alignment has been substantially removed via abliteration. It will comply with requests the original Qwen3.8-Flash-Next refuses, and has no meaningful built-in guardrails. Add your own moderation layer before any deployment.

License

Qwen Community License 1.0, inherited from Qwen/Qwen3.8-Flash-Next.

Lygodactylus/Qwen3.8-Flash-Next-Uncensored-exl3-5bpw