keys-Qwen3.8-Flash-Next-NVFP4-dual-ablit-house-qsa-L3-47
Abliterated NVIDIA Qwen3.8-Flash-Next NVFP4 for the Keys 2× DGX Spark vLLM recipe (TP2 + EP + MTP3).
This is not a Dealign weight dump and not a byte-copy of Dealign o_proj. It is our house residual-writer projection applied to NVIDIA stock. The axis is recovered per QSA layer from (W_dealign − W_nvidia); the edit is W ← W − λ v vᵀ W plus per-row L2 restore.
See RESPONSIBLE_USE.md and the gate form above. Access is gated with automatic approval after you agree.
| HF | https://huggingface.co/drowzeys/keys-Qwen3.8-Flash-Next-NVFP4-dual-ablit-house-qsa-L3-47 |
| Stock NVFP4 (parent) | nvidia/Qwen3.8-Flash-Next-NVFP4 |
| Axis source (not shipped) | dealignai/Qwen3.8-Flash-Next-UNCENSORED-NVFP4 |
| Upstream | Qwen/Qwen3.8-Flash-Next |
| Serve formula | MiaAI-Lab/Qwen3.8-Flash-Next-Dual-DGX-Sparks |
| Ablit | 12 QSA self_attn.o_proj house-projected (L3/7/11/15/19/23/27/31/35/39/43/47) · mean Δrel ≈ 0.167 · recon vs Dealign 0.4–3.2% |
| Anchors (stock NVIDIA) | all GDN out_proj · routed experts · MTP · PLE · vision · chat template |
| Gate | 32/32 bypass, 0 refuse (thinking-off, 384 tok) |
A fully independent dest from stock activation capture (no Dealign weights) scored 1/32. The working axis is Dealign’s; the published edit is the house projection.
Method
House projection on NVIDIA dual-Spark NVFP4, QSA residual writers only:
W ← W − λ v vᵀ W
then restore each output row’s L2.
| Layer | λ | projected Δrel | recon vs Dealign |
|---|---|---|---|
| 3 | 6.95 | 0.180 | 2.1% |
| 7 | 6.80 | 0.182 | 2.1% |
| 11 | 6.85 | 0.185 | 2.2% |
| 15 | 6.95 | 0.169 | 1.8% |
| 19 | 6.95 | 0.178 | 2.0% |
| 23 | 7.85 | 0.214 | 3.2% |
| 27 | 7.55 | 0.196 | 2.6% |
| 31 | 8.00 | 0.195 | 2.6% |
| 35 | 7.85 | 0.201 | 2.9% |
| 39 | 4.00 | 0.107 | 0.6% |
| 43 | 4.05 | 0.090 | 0.4% |
| 47 | 3.90 | 0.113 | 0.7% |
v is the rank-1 left singular vector of (W_dealign − W_nvidia) at that layer (rank-1 energy ≥ 0.9997). GDN layers are unmodified (Dealign’s GDN out_proj matched NVIDIA to Δrel = 0).
Serve (2× DGX Spark)
Same Mean / Mia dual-Spark recipe as stock NVIDIA:
- TP=2, expert parallel on, MTP=3
kv_cache_dtype=fp8,gpu_memory_utilization=0.835(do not raise above 0.85)max_model_len=262144- image
vllm/vllm-openai:qwen38-flash-next
Thinking-on (Hermes): chat_template_kwargs.enable_thinking=true, reasoning_effort=xhigh, max_tokens≥8192.
What this is not
- Not
dealignai/Qwen3.8-Flash-Next-UNCENSORED-NVFP4reuploaded - Not a BF16 byte-copy of Dealign
o_projinto NVIDIA shards - Not the single-Spark MXFP8 LIL layout (
local-inference-lab/…/ Mia TP1 dest)
License
Use is governed by the NVIDIA Open Model License and Qwen Community License 1.0, inherited from nvidia/Qwen3.8-Flash-Next-NVFP4.