license: apache-2.0
base_model: huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated
base_model_relation: quantized
library_name: mlx
pipeline_tag: text-generation
tags:
- mlx
- apple-silicon
- uncensored
- abliterated
- coder
- qwen3_5
- hybrid-attention
- 8bit
language: - en
- zh
Huihui Qwen3.8 27B Coder390 EfficientThink abliterated — MLX 8bit (Apple Silicon)
NovaMLX-format 8-bit build of the uncensored (huihui-ai abliterated)
Qwen3.8-27B-Coder390-EfficientThink coder: a strong 27.8B hybrid-attention coder
with the refusal direction removed. Runs on any Apple Silicon Mac with 48GB+
unified memory at ~3× the speed of the BF16 build.
⚠️ Uncensored model. Refusal behaviour was removed upstream by abliteration.
You are responsible for how you use it and for complying with applicable law.
Lineage
- nerkyor/Qwen3.8-27B-Coder390-EfficientThink-... — code/reasoning fine-tune
of Qwen3.8-27B with MTP head. - huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated —
abliterated (decensored) by huihui-ai. - hakunokina/...-oQ8e-mtp — oMLX oQ8e quantization (imatrix-calibrated 8-bit
affine, group 64), KL 0.00115 / top-1 98.6% vs BF16. - This repo — converted to standard MLX tensor naming by
NovaMLX so it loads in any MLX runtime
(the oMLX build'slanguage_model.*layout only loads in oMLX ≥ 0.7.1).
MTP head kept inmtp.safetensors.
Specs
| Architecture | Qwen3_5ForConditionalGeneration (text-only weights) |
| Layers | 64 (48 linear + 16 full attention) |
| Quantization | 8-bit affine, group 64, imatrix-calibrated (via oQ8e) |
| Size | ~29 GB |
| Context | 262,144 tokens |
Benchmarks (oMLX accuracy suite, from the oQ8e build, thinking off)
| Benchmark | Score |
|---|---|
| MMLU-Pro | 64.7 |
| GSM8K | 95.0 |
| HumanEval | 90.8 |
| MBPP | 80.0 |
| LiveCodeBench | 68.0 |
Speed (Mac Studio M4 Max, 128 GB, NovaMLX server)
| BF16 build | ~5.4 tok/s |
| This build (8-bit) | ~15 tok/s decode |
| Memory resident | ~27 GB wired |
Credits
- nerkyor — Coder390 EfficientThink training
- huihui-ai — abliteration
- hakunokina — oQ8e quantization
- NovaMLX (数字流动) — format conversion & verification