DogEasy/Qwen3.8-Flash-Next-Abliterated-HGNv1

Qwen3.8-Flash-Next — Abliterated · HGN-HQ (λ=1.5)

HGN-HQ quantized build of windowsxp811203/Qwen3.8-Flash-Next-Abliterated for halogen-flash-server and the gfx1151-engine runtime.

This is not a transformers / vLLM / llama.cpp checkpoint. There are no safetensors or GGUF files here; the .hgn set only loads in an HGN-compatible engine. For the bf16 weights, use the source repo above.

What this build is

The source checkpoint removes refusal by orthogonalizing 151 residual-stream-writing tensors against a captured direction at λ=1.5 (W_ablit = (I − λ r rᵀ) W_stock). Naively quantizing those edited tensors would blur exactly the edit the model exists for — so this build quantizes around the direction and then repairs the experts against it:

1. Direction recovery. The shared refusal direction r is recovered from the official Halogen quality overlay (qwen38-flash-next-w4b.overlay.hgn) and the abliterated BF16 checkpoint, using 12 anchor tensors (all self_attn.o_proj / linear_attn.out_proj / shared_expert.down_proj spread across the 48 layers). Recovery was effectively exact: min anchor cosine 0.99995, median s1/s2 ≈ 57, per-anchor λ estimates all within 1.4994–1.5001 of the configured λ=1.5. Full audit in Qwen3.8-Flash-Next-Abliterated.halogen.overlay.hgn.manifest.json.
2. HGN-HQ conversion. Current upstream flashnext2hgn pipeline: routed experts → weighted Q4CP with the universal expert codebook (x² weighting, no imatrix), dense tensors → q8g32 overlay, MTP sidecar → q8g64, PLE table → FP8 E4M3 + global scale, vision tower → BF16. Few-row dense exceptions stay BF16 per upstream policy.
3. Direction-preserving expert repair. All 49 fused *.mlp.experts.down_proj.weight tensors (48 trunk + 1 MTP) are patched nibble-by-nibble so that rᵀQ ≈ rᵀW_ablit — the repair only flips adjacent Q4 levels when that reduces directional error. It does not push the projection toward zero, so it does not silently turn λ=1.5 into λ≈1.

Verification (shipped artifacts, measured)

From Qwen3.8-Flash-Next-Abliterated.experts.json:

metric	value	gate
repaired expert tensors	49	48 trunk + 1 MTP
max direction residual ratio	0.0201	≤ 0.05 (hard fail)
mean direction residual ratio	0.0200	—
weighted error regressions vs absmax baseline	0	0 (fail-closed restore)
max nibble flip fraction	0.154 %	—

Any tensor the repair would have made worse than the same-codebook absmax baseline was restored to the upstream quantized version and the run failed closed; zero tensors hit that path.

No downstream benchmark re-run was performed on the quantized weights — the capability and refusal numbers on the source model card are properties of the bf16 build. What is verified here is that the λ=1.5 direction edit survives quantization to within a 2 % residual on every fused expert tensor.

Files

file	size	role
Qwen3.8-Flash-Next-Abliterated.hgn	115.5 GiB	base weights (routed experts Q4CP, dense base)
Qwen3.8-Flash-Next-Abliterated.overlay.hgn	4.83 GiB	production main overlay — q8g32 HGN-HQ dense
Qwen3.8-Flash-Next-Abliterated-mtp.hgn	90 MiB	MTP draft head sidecar (q8g64)
Qwen3.8-Flash-Next-Abliterated-vision.hgn	858 MiB	vision tower (BF16)
Qwen3.8-Flash-Next-Abliterated.halogen.overlay.hgn	2.31 GiB	phase-1 audit artifact — do not load in production
…halogen.overlay.hgn.manifest.json / …report.tsv	—	direction-recovery audit data
Qwen3.8-Flash-Next-Abliterated.direction.npy	10 KiB	recovered refusal direction r, (2560,) float32, ‖r‖≈1
Qwen3.8-Flash-Next-Abliterated.experts.json	730 KiB	per-tensor expert repair report
tokenizer/	—	tokenizer, chat template, processor configs

Load order matters:

base .hgn  →  .overlay.hgn (q8g32 main)  →  -mtp.hgn  →  -vision.hgn

Do not additionally load .halogen.overlay.hgn after the q8g32 main overlay — it is a legacy calibration-transfer artifact and would overwrite the current 8-bit dense tensors with 4-bit Halogen-derived ones, defeating the HGN-HQ route.

Runtime

• halogen-flash-server or gfx1151-engine, with HGN-HQ (q8g32 overlay) support.
• Converted and validated on a Ryzen AI Max+ 395 (gfx1151) with the TheRock / ROCm 10.0.0 stable stack; on that machine the 10.1/10.2 stacks fail a rocBLAS SGEMM M=640,N=1,K=2560 probe, so 10.0.0 is the known-good runtime.
• On startup the engine should report dense: 8-bit (q8g32) weights present, few-row bf16 gemv on.

License

Qwen Community License 1.0 — inherited from the base model through the abliterated checkpoint; see the source repo's LICENSE. Redistribution and derivative weights are permitted. Two conditions to read before deploying:

• Products above 100 M MAU or US$20 M monthly revenue must display the model name prominently.
• Running a Model-as-a-Service or AI Work Assistant business requires a separate license from Qwen before commercial use. Internal use is exempt provided the model, its outputs and its capabilities are not made available to third parties.

Support / 打賞

If these models are useful to you, tips are appreciated — they pay for the GPU time. 如果這些模型對你有幫助，歡迎打賞，用於支應算力成本。

USDT (TRC20) · TPTo32r7vKazpTNaFqfFZ2ztoK1DG88888

Disclaimer

This model will not refuse. It is published for alignment and safety research. You are responsible for your use of it and for complying with applicable law.