Qwen3.8-Flash-Next-NVFP4-ablit-a070
A milder abliteration of Qwen3.8-Flash-Next for one DGX Spark (GB10, TP=1) in the Mia single-Spark NVFP4 recipe layout. It is a linear interpolation between the stock Mia checkpoint and Keys' abliterated splice, at strength α = 0.7:
W_new = W_stock + 0.7 · (W_ablit − W_stock)
applied to the nine attention o_proj tensors that Keys' ablit edits. Everything else is
byte-identical to stock.
Personal build. Not affiliated with Qwen, Mia-AiLab, local-inference-lab or Keys.
Why this exists
Keys' full-strength ablit (drowzeys/keys-Qwen3.8-flash-next-ablit-Mia-Single-Spark-only)
removes refusals, but on this host it degenerates exactly in the region an ablit is used for.
Neutral prompts are fine. Any input that stock would refuse or deflect collapses into empty
numbered lists (1. Hello, 1.) and stays there for the rest of the conversation. The serving
profile is not the cause: fp8 KV cache, bf16 SSM state and the packed PLE table were all ruled
out by measurement. The effect is in the weights.
Analysis of the delta per edited layer shows an almost exactly rank-1 injection: energy along the delta direction rises about 6× relative to stock (|uᵀW| 2.1 → 12.2), and the nine directions are nearly uncorrelated across layers. Scaling that injection down keeps the refusal bypass while restoring coherence in the boundary region.
What was changed
| Base | Mia-AiLab/Qwen3.8-Flash-Next-NVFP4, snapshot 925d7be6c14c6c9442ef83e8f05b5a3c39304f69 |
| Ablit source | drowzeys/keys-Qwen3.8-flash-next-ablit-Mia-Single-Spark-only, snapshot b3b60a4ce9d5922e8395887a1e5a58734b51b522 |
| Edited tensors | model.language_model.layers.{15,19,23,27,31,35,39,43,47}.self_attn.o_proj.weight + weight_scale |
| Shards rewritten | 9 of 34 (model-00023, 00025–00030, 00032, 00033) |
| Method | dequantize both MXFP8 tensors (F8_E4M3 codes × e8m0 group-32 scales), interpolate at α = 0.7, requantize with per-group scale = max(stock, ablit), round-to-nearest F8_E4M3 |
| Requant quality | cosine ≥ 0.9998 to the exact interpolated tensor on every layer |
| Effective α per layer | 0.704–0.746 (rounding biases slightly toward the ablit; see MILD_ABLIT_META.json) |
| Unchanged | MTP head, PLE n-gram table, routed experts, GDN out_proj, vision, layers 0–14, config, tokenizer, chat template, index |
MILD_ABLIT_META.json records the method, the source snapshots and the effective α for each layer.
Strength comparison
| Variant | Neutral prompts | Hostile input | Edgy content on request | Notes |
|---|---|---|---|---|
| Keys ablit (α = 1.0) | clean | collapses to 1. Hello / 1. |
empty numbered lists | refusal-adjacent region dead |
| α = 0.6 | clean | coherent meta-reply | coherent prose | none |
| α = 0.7 (this repo) | clean | coherent meta-reply | most coherent of the series | none |
| α = 0.8 | clean | coherent but non-sequitur | cruder, but back to list form | occasionally leaks an internal analysis scaffold as a numbered list |
α only scales the strength of the injected anti-refusal direction. Higher α means fewer refusals, but from about 0.8 the list tic creeps back. Higher α never makes the model smarter or more creative; the capabilities live in stock, and the better output quality at 0.6–0.7 is recovered coherence, not added ability.
Not measured: bypass on hard refusals; the comparison above covers mild cases only. If a real refusal shows up at 0.7, try 0.8 and re-check coherence.
Compatibility
This is the Mia single-Spark layout (34 shards, MXFP8 attention o_proj, NVFP4 routed
experts, packed PLE). It is not a drop-in for any other Qwen3.8-Flash-Next quant (NVIDIA
ModelOpt / RadixArk experts-only NVFP4, Inferact, official BF16/FP8, GGUF, MLX): those store
o_proj in a different dtype and shard map. Do not copy these shards onto another tree.
Serve with the Mia single-Spark vLLM recipe (vllm/vllm-openai:qwen38-flash-next, TP=1,
in-checkpoint MTP D=3, packed PLE table). The memory footprint is identical to stock and to
Keys' ablit; the recipe's shipped profile (KV_CACHE_DTYPE=fp8, MAMBA_SSM_CACHE_DTYPE=bfloat16)
works unchanged. The packed PLE table can be reused from stock, the PLE shards are untouched.
Responsible use
Access is gated. By requesting access, downloading, or using these weights you agree to the terms shown on the access form (they are the same terms as the source ablit). Safety refusals are weakened in this checkpoint. Filtering, human review and access control are yours to supply. The ablit delta originates from a gated checkpoint whose access terms you accepted; those terms carry over to this derivative: no sexual content involving minors, no material promoting self-harm or suicide, no harassment, doxxing or fraud targeting real people, nothing illegal in your jurisdiction, and nothing barred by the upstream Qwen Community License. Users must be 18 or older. The weights are provided as-is, without warranty.
Credits
- Qwen / Alibaba — Qwen3.8-Flash-Next, the base model
- MiaAI-Lab — the single-DGX-Spark vLLM recipe and
Mia-AiLab/Qwen3.8-Flash-Next-NVFP4, the stock checkpoint every unchanged byte here comes from - local-inference-lab — the original NVFP4 / ModelOpt quantization that Mia's repo mirrors:
local-inference-lab/Qwen3.8-Flash-Next-NVFP4 - Keys (drowzeys) — the abliteration splice this build interpolates toward:
drowzeys/keys-Qwen3.8-flash-next-ablit-Mia-Single-Spark-only