license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
library_name: vllm
pipeline_tag: image-text-to-text
tags:
- qwen3_5
- qwen3.8
- abliterated
- uncensored
- huihui
- compressed-tensors
- w8a16
- int8
- quantized
- vllm
- mtp
- speculative-decoding
- vision
- conversational
base_model_relation: quantized
Qwen3.8-27B Abliterated · INT8 W8A16 · BF16 MTP
An abliterated, Ampere-optimized quantization for dual RTX 3090 inference.
Base model · Qwen3.8-27B · remove-refusals-with-transformers · vLLM · llm-compressor
[!NOTE]
This is a numerical W8A16 quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated, an abliterated (uncensored) version of Qwen/Qwen3.8-27B. All model credit belongs to huihui-ai (the abliteration) and Qwen (the base model); refer to the upstream model card for architecture, capabilities, and usage guidance. This repository changes numerics only.
[!TIP]
This build targets vLLM on Ampere-class GPUs where native FP8 tensor-core execution is unavailable. It retains the model's vision tower, thinking controls, native context, and the byte-preserved BF16 MTP head.
Relationship to the sibling quantization
This checkpoint is the abliterated counterpart of lued/Qwen3.8-27B-INT8-W8A16-MTP. The repositories share:
- the identical W8A16
compressed-tensorsrecipe (400 packed GEMMs, same BF16 preservation set, group-128 symmetric data-free RTN); - the identical native BF16 MTP head (15 tensors, byte-identical to Qwen's original MTP — same sha256 as the sibling release);
- the identical tokenizer, chat template, processor configs, and bounded-shard layout;
- the same total checkpoint footprint (29.44 GiB).
The only deviation is the base language-model weights: this checkpoint's LM weights come from huihui-ai's abliterated BF16 model instead of stock Qwen3.8-27B.
Abliteration (inherited from the base)
Per the upstream card, the source is an uncensored version of Qwen/Qwen3.8-27B created with remove-refusals-with-transformers — a crude, proof-of-concept refusal-removal technique that does not use TransformerLens. The first 15 layers were retained without ablation; MTP and visual components were not modified. These are the upstream card's claims, not measurements made here. The W8A16 step touches only stored numerics of the 400 GEMMs and does not change the refusal profile.
Quantization fidelity
Measured on this exact checkpoint against its own BF16 teacher (the huihui source) on 2026-08-17 with tools/measure_kld.py (teacher-forced, full-vocabulary FP32 log-softmax, FP64 accumulation, nats/token; PyTorch 2.11.0+cu128, Transformers 5.10.1, compressed-tensors 0.17.1):
| Candidate | Mean KLD vs BF16 | Top-1 agreement | Weight files |
|---|---|---|---|
| This W8 / BF16 A16 | 0.000705 | 98.72% | 29.44 GiB |
The KLD result measures checkpoint-weight drift over 467 teacher-forced positions (fixed 8-prompt short suite). It is not a functional quality score; tool use, coding, multimodal quality, and long-context recall require separate behavioral evaluation.
Short-prompt suite
| Prompt | Category | Tokens | Positions | Mean KLD | Top-1 agreement |
|---|---|---|---|---|---|
factual |
factual | 13 | 12 | 0.000415 | 1.0000 |
code-fib |
code | 87 | 86 | 0.000606 | 0.9884 |
physics-uncertainty |
physics | 86 | 85 | 0.000826 | 0.9882 |
math-train |
math | 58 | 57 | 0.000637 | 1.0000 |
sql-top5 |
technical | 60 | 59 | 0.000894 | 0.9831 |
narrative-clock |
narrative | 68 | 67 | 0.000784 | 1.0000 |
instruction-stack |
instruction | 43 | 42 | 0.000589 | 0.9762 |
history-industrial |
history | 60 | 59 | 0.000606 | 0.9661 |
| Token-weighted mean | 467 | 0.000705 |
kld_result.json in this repository is the single source of truth for these numbers.
Checkpoint profile
| Property | Value |
|---|---|
| Quantization | Data-free symmetric RTN W8A16, group size 128 |
| Runtime format | compressed-tensors / pack-quantized |
| Quantized GEMMs | 400: 192 MLP, 64 full-attention, 144 dense GDN projections |
| Preserved precision | BF16 vision tower, lm_head, MTP (15 tensors, byte-identical to Qwen's original MTP), and recurrent GDN gates (96) |
| Structural audit | STRUCTURAL AUDIT PASS (packed set, ignore list, index contract) |
| MTP audit | MTP AUDIT PASS (15 tensors dtype-, shape-, and bit-identical to the source) |
| Shard layout | 7 main shards (5.4 GiB max) + model_mtp.safetensors (849,400,392 bytes) |
| Runtime | vLLM; this repository is not a GGUF checkpoint |
MTP preservation
The MTP head is the exact BF16 MTP from the original Qwen3.8-27B — never touched by the abliteration and never re-quantized:
- 15 tensors in
model_mtp.safetensors(849,400,392 bytes). - sha256
1d8268aa85ace093a561e3e7b63b9d390dac1cd55a90cd55b5ec509c3c9da9fe— identical to the MTP bytes in the Qwen3.8-27B source, in the AEON source, and in the sibling release lued/Qwen3.8-27B-INT8-W8A16-MTP. - All 15 tensors verified
bfloat16, shape-equal, and bit-equal to the huihui source (strict audit, not a count check).
Reproduction contract
The immutable source is huihui-ai/Huihui-Qwen3.8-27B-abliterated commit d42ca8978c5a66e92c3446d46e8adfe03ef692ff. The exact selective W8A16 configuration is recorded in recipe.yaml and config.json. A conforming rebuild must take native MTP from that source, reproduce the source-derived packed projection set, and prove all 15 MTP tensors bit-identical before resharing for publication.
Status
Built on 2026-08-17: both release audits PASS, KLD measured, published tolued/Qwen3.8-27B-huihui-abliterated-INT8-W8A16-MTP.
Serving validated on 2026-08-18: boots through llama-swap from the lued/*
cache entry and serves a smoke request.