Huihui-Qwen3.8-27B-abliterated-NVFP4

Huihui-Qwen3.8-27B-abliterated-NVFP4

Mixed-precision NVFP4 quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated, built with llm-compressor.

24.7 GB. Calibrated on text generated by this abliterated model itself, not by stock Qwen — see below, it matters.

Recipe

AWQ per-input-channel scaling, then AutoRound (SignSGD, block-wise loss, 200 iters) on the NVFP4 MLPs and GPTQ on the 8-bit modules. Requires Blackwell for native NVFP4.

Calibration: self-distilled from the abliterated model on a balanced Nemotron-v2 prompt blend (25% code, 25% math, 20% STEM, 20% chat, 10% multilingual).

Benchmarks

Measured against the abliterated BF16 model as its own reference — not stock Qwen — so the numbers reflect quantization damage only, not the effect of abliteration. 142,727 tokens plus 200 free greedy generations. vLLM 0.27.1, TP=2, 2×B300.

Sizes are on-disk tensor bytes and include the ~0.85 GB BF16 MTP head.

Columns. top-1 is raw argmax agreement with the BF16 abliterated model. The bucket columns are disagreement rates split by how confident the reference was at that position (top1−top2 logprob margin): near-tie <0.5, moderate 0.5–2, confident 2–5, certain >5. Only confident and certain are real damage. divmed is the median token index at which free greedy generation first diverges.

Perplexity is excluded — on this model family it is anti-correlated with quality.

Calibration matters more than abliteration

An earlier build of this model used the stock-Qwen calibration set and a weaker recipe, and landed at 3.54% confident damage. Regenerating the calibration from the abliterated model itself brings that to 1.80%.

That also answers a question worth stating plainly: abliterated weights are not intrinsically harder to quantize. With matched recipe and self-distilled calibration this model reaches 1.80% confident damage, against 1.85% for the same recipe on stock Qwen3.8-27B. The earlier gap was the calibration and recipe, not the abliteration.

NVFP4 vs INT4 on this model

The INT4 sibling is more faithful (confident 0.88% vs 1.80%) but decodes at 4551 tok/s against 10702 here. This build is the throughput choice on Blackwell; the INT4 one is the fidelity choice, and the only option on Ampere/Ada where FP4 does not exist.

Usage

from vllm import LLM
llm = LLM("TelperionAI/Huihui-Qwen3.8-27B-abliterated-NVFP4", tensor_parallel_size=2)

Speculative decoding (MTP)

The MTP head is included, in BF16, grafted from the abliterated base (not stock Qwen):

llm = LLM("TelperionAI/Huihui-Qwen3.8-27B-abliterated-NVFP4", tensor_parallel_size=2,
          speculative_config={"method": "mtp", "num_speculative_tokens": 2})

Qwen3_5ForConditionalGeneration does not carry mtp.* in its state dict, so llm-compressor silently drops it even though config.json declares mtp_num_hidden_layers: 1. It is excluded from quantization via re:.*mtp.*. Acceptance rate has not been measured; the head is verified to load and generate.

Limitations

• Single evaluation corpus, and no downstream task benchmarks.
• Abliterated base. This model has had its refusal directions removed upstream; that behaviour is inherited here and is not something quantization changes.
• The abliterated calibration set is ~18% smaller than the stock one (the same length filter kept fewer generations), so it is not perfectly matched to the stock-model builds.
• Vision tower untouched (BF16); evaluated as a text model.
• MTP acceptance rate unmeasured.