license: apache-2.0
library_name: safetensors
base_model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
pipeline_tag: image-text-to-text
tags:
- candle
- qwen3_5
- vision
- sceneworks
Huihui-Qwen3.8-27B-abliterated — SceneWorks Candle rehost (q8 / q4)
Pre-quantized Candle tiers ofhuihui-ai/Huihui-Qwen3.8-27B-abliterated
at revision 739e3c5b89849f6c238ce1e5b70008612ae42cdd, an abliterated build ofQwen/Qwen3.8-27B. Published by SceneWorks so the
tiers resolve by an immutable commit SHA for SceneWorks Inference's Candle language-model engine,
and so a q4 user does not download the 55.6 GB full-precision files. It is not an official
Qwen or huihui-ai distribution. The MLX tiers are inSceneWorks/qwen3.8-27b-abliterated-mlx.
Layout
| Dir | Contents |
|---|---|
q8/ |
pre-quantized Q8 tier (32.4 GB): projections stored as GGML Q8_0 blocks |
q4/ |
pre-quantized Q4 tier (20.0 GB): projections stored as GGML Q4_K blocks |
Each tier directory is self-contained (config, tokenizer, chat template, weights). Download only
the tier you use. Full precision is not mirrored here; take it from the upstream repository.
The tiers are produced from the upstream bf16 snapshot by candle-llm's snapshot preparer
(prepare_snapshot). Each decoder and MTP projection is quantized once and stored as its raw GGML
blocks, a U8 tensor [rows, blocks_per_row, block_bytes], which the Candle loader rebuilds
directly (no dequantize/re-quantize). config.json carries quantization: {bits, storage: "ggml"}.
The vision tower (model.visual.*), embeddings, the LM head and norms stay bf16 in every tier, so
the model still accepts image input.
Verification status
q4/: loaded by candle-llm on Apple Metal and grounded on an image (named a solid red and a
solid blue image correctly).q8/: written by the same preparer; not yet loaded on an accelerator (the test host refused
the load for memory).- Neither tier has been run on CUDA yet.
Provenance
SOURCE_REVISION.json records the upstream repository and revision, the base model, and the
SceneWorks Inference revision that produced the tiers.
License and use
Apache-2.0, as released by Qwen and kept by huihui-ai; see LICENSE. The upstream model card
states that abliteration removes the model's refusal behaviour and that its output may be
inappropriate for some audiences. That applies to these tiers unchanged. You are responsible
for how you use it.