NVIDIA ModelOpt quantization of the Darkstar Abliterated BF16 derivative of Qwen/Qwen3.8-27B, pinned upstream at 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The R3 edit projects a normalized float32 refusal direction from exactly 131 residual-writing tensors at layer 38 with seed 42. The selected mixed quantization uses W4A16 NVFP4 group 16 for language MLP projections and lm_head, FP8 e4m3 for self-attention and GatedDeltaNet projections, and BF16 for protected components and runtime KV.

ModelOpt is pinned to 0.46.0rc2 at 43fd41a58d52c4e6e5dec1d1ff5989ecc737ae1a; the selected operator recipe SHA-256 is 90fc6b37c00334debd49f1975ab406b5e20667f07e4be0be3e463a648abac642. Calibration used cnn_dailymail plus nemotron-post-training-dataset-v2, 512+512 samples, sequence length 2048, seed 1234. All 15 source BF16 MTP tensors were preserved. Artifact identity: _SUCCESS.json SHA-256 3d89ec57c1371e142adc2584de079b54a0e1d8c12dc9550118d0a851da020a79; manifest SHA-256 642dbbe89b085a2daf5119c37c0496576a475ed64c36653fc993c04abaf2ca9f.

Full provenance, protocol, and caveats:

Evaluation

Metric	Value	Basis
GPQA Diamond (thinking off, matched)	148/198 = 74.75%	198/198 terminal parseable; 0 timeout/parse/error
GPQA Diamond (thinking on, secondary)	164/198 = 82.83%	rejected historical compressed-tensors artifact; not attributable to this ModelOpt build
Quantization delta vs Abliterated BF16	+2 questions / +1.01 pp	146/198 → 148/198
Harmful-prompt compliance	200/200 (0/200 refusals)	283/283 terminal; 0 errors
Safe over-refusals	0/83 (0.00%)	0 errors
Single-stream throughput (MTP10)	mean 251.889 tok/s	nonmonotonic MTP1-12 sweep; MTP8 mean 250.862 tok/s

Matched GPQA evidence SHA-256: summary d8d0b5c0de686846338ce89e9a55456baec0550bbad765ccc65e9fa57380b818, journal 9bb4913202977bad204ebde8d2e31e8357a3308f3b77c44539ef3977a2c6e813. Behavior evidence SHA-256: summary d814eac6eef86cb32c891d5c3b1765be806cb0fb634173080cd5df46ea9f9233, journal 7b6ddf556ab3afc1f8582041d7b723dbfeaeb24b02b8bd27562b0c9928a37d4f.

Serving-capacity measurements at the frozen MTP10 profile used 512 generated tokens, two repeats, zero failed requests, and zero fatal markers:

Prompt	Prompt tokens	C1 mean aggregate tok/s (pass 1 / pass 4)	C2 mean aggregate tok/s (pass 1 / pass 4)
4K chars	738	193.411 / 195.486	352.313 / 356.564
16K chars	2653	183.611 / 183.134	328.393 / 343.799
48K chars	7758	157.579 / 157.889	286.029 / 284.465

Safety warning

This model has had its refusal direction deliberately reduced and has no added safety mitigations. It complied with 200/200 harmful prompts in the measured suite. Deploy only behind appropriate policy, filtering, access controls, and legal review. Refusal-rate numbers are behavior measurements, not safety endorsements.

Release reference

Engineering release: darkstar-qwen3.8-27b-v1.0.0. This immutable tag exists and the release contract is published.

Runtime

Validated with vLLM 0.27.1, compiled mode, Flash Attention, BF16 KV cache, context 126,144, MTP depth 10, 32K scheduler budget, max_num_seqs=16, prefix caching, and chunked prefill. MTP10 was selected on mean throughput after nonmonotonic confirmation:

bash

VLLM_ATTENTION_BACKEND=FLASH_ATTN \
vllm serve HangGlidersRule/Darkstar-Qwen3.8-27B-Abliterated-ModelOpt-W4A16-NVFP4-Mixed-FP8 \
  --served-model-name darkstar-qwen38-abliterated-nvfp4 \
  --kv-cache-dtype bf16 \
  --max-model-len 126144 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 32768 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --compilation-config 2 \
  --speculative-config '{"method": "mtp", "num_speculative_tokens": 10}'