license: other
license_name: qwen-community-license-1.0
license_link: LICENSE
base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored
base_model_relation: quantized
library_name: exllamav3
tags:

• exl3
• moe
• uncensored
• qwen3.8
• flash-next

Qwen3.8-Flash-Next-Uncensored: EXL3, shared input vector, 4.03 bpw

An EXL3 quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored
(commit d9dbbb7f49cad93a0e656ea69200f6b58600714a), made 2026-10-20.

How it was made

• exllamav3 1.5.2 with turboderp's 4.05 settings (-b 4.05 -hb 6 -mb 4 --out_scales always -cb mul1, default
  calibration 250 x 2048), plus a per-tensor recipe that puts every routed expert at 4 bpw (the budgeted allocator
  would put layer 0's experts at 5). Attention 6 bpw; delta-net and shared experts 5-6 bpw, as turboderp's 4.05
  allocation; lm_head 6, vision 6, MTP head 4. Overall 4.03 bpw.
• Shared input vector: within each layer, all routed gate/up matrices share one input sign/scale vector (suh),
  followed by a joint refit of the scales over the layer. This is for the shared-input-vector kernels of
  radiance-exl3 (in development). The tensors are in the standard EXL3
  format; this repo was not tested in exllamav3 itself.
• The n-gram table ngram_embedding.safetensors is turboderp's, from
  turboderp/Qwen3.8-Flash-Next-exl3@55a732e0c4c3d4614bc42b68493bb930d9b02c0a. It applies unchanged: all 128 n-gram
  tensors of orcarouter's model are byte-identical to base Qwen3.8-Flash-Next's (checked).
• Conversion log, progress log and the first-layer check are in logs/. ALL_FILES.sha256 lists every file's sha256
  as written by the converter.

Quality was not measured for this model. On base Qwen3.8-Flash-Next the same method (with turboderp's layer 0)
scored KL 0.0486 against bf16 vs 0.0467 for turboderp's 4.05 (difference not significant; see
Dyluhn/Qwen3.8-Flash-Next-exl3-sharedvec-radiance).

The original model card of orcarouter's model is kept as ORIGINAL_MODEL_CARD.md. Licence: Qwen Community License 1.0 (LICENSE).