Qwen3.8-Flash-Next AutoRound Hybrid — Uncensored

Abliterated (refusal-removed) build of
Saren/Qwen3.8-Flash-Next-W4A16-AutoRound-hybrid
— Qwen3.8-Flash-Next (125B MoE + 51B n-gram table, 6B active) quantized as
int4 MoE (W4A16 AutoRound/GPTQ-Marlin g128) + int8 lm_head + blockwise-fp8
side layers. Same weights format, same serving stack, refusals removed.

Method

No fine-tuning, no requantization pipeline. Weight-level edit only:

1. Recovered the refusal direction from
   orcarouter/Qwen3.8-Flash-Next-Uncensored: a stock-vs-abliterated shard
   diff is cleanly rank-1 (≈100:1) with the identical vector (cos = 1.0000)
   across attention o_proj, GDN out_proj, and shared-expert down_proj, while
   gates/inputs/embeddings are untouched.
2. Applied W − r(rᵀW) to all 25,185 residual-writing matrices in-quant:
   96 fp8 (12 QSA o_proj + 36 GDN out_proj + 48 shared-expert down_proj,
   blockwise-e4m3 requant), 24,576 int4 MoE-expert down_proj (fresh
   per-column symmetric scales — fixed scales would round a 2% edit back to
   ~zero on 4-bit levels), 513 bf16 MTP draft-layer tensors. Zero skips.
3. PLE n-gram table unchanged — reuse
   Saren/Qwen3.8-Flash-Next-ple-table-fp8
   via mmap.

Measured (1× DGX Spark, vLLM + MTP=3, same config as stock hybrid)

• 5/5 refusal probes comply (no refusal phrases), coherent completions
• Capability intact: codegen w/ tests, long-form essay, math reasoning
• Decode ~32 tok/s, TTFT ~230 ms — no regression vs stock hybrid

Serve it

git clone <installer-repo> && cd <installer-repo>
docker build -t qwen38-flash-dgx <saren-recipe>  # serving image w/ patches
MODEL_ID=bidhata/Qwen3.8-Flash-Next-AutoRound-Uncensored ./serve.sh

Needs the PLE table dir alongside (TABLE_ID=Saren/Qwen3.8-Flash-Next-ple-table-fp8,
auto-downloaded by serve.sh), ~130 GB disk, 128 GB unified memory.

Credits / license

• Weights: Qwen team (Qwen3.8-Flash-Next) → Intel (W4A16 AutoRound) →
  Saren-Arterius (hybrid) → refusal-direction transfer from orcarouter's
  abliteration.
• Carries the Qwen community license — review before production use.
• Uncensored research artifact. You are solely responsible for what you
  generate and for complying with applicable law.

Original build notes: see README.build.md.