license: apache-2.0
base_model: xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
base_model_relation: quantized
library_name: mlx
pipeline_tag: text-generation
tags:
- mlx
- apple-silicon
- uncensored
- qwen3_5
- hybrid-attention
- linear-attention
- distilled
- 4bit
language: - en
- zh
DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED MLX 4bit (Apple Silicon)
First MLX build of the viral DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic —
DeepSeek-V4-Pro reasoning distilled into Qwen3.5-9B, then decensored with the
Heretic tool's Arbitrary-Rank Ablation. Runs on any Apple Silicon Mac with 16GB+
unified memory.
⚠️ Uncensored model. Refusal behaviour was removed upstream by abliteration
(refusals 100/100 → 3/100, KL 0.042 from base). You are responsible for how you
use it and for complying with applicable law.
Lineage
- Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B — knowledge distillation from
DeepSeek-V4-Pro on ~250K math/STEM samples (SFT + GSPO RL), Qwen3.5-9B backbone. - xtools-at/...-UNCENSORED-heretic — decensored via Heretic v2 ARA, layers 9–24.
- This repo — quantized to MLX 4-bit affine (group 64) from the BF16 source
by NovaMLX — 数字流动 (Shenzhen Digitalflow Software).
Why this build
- First and only MLX conversion of this model (previously GGUF/NVFP4 only —
unusable in MLX runtimes). - Hybrid attention (75% GatedDeltaNet linear attention + 25% full attention):
fast on Apple GPU, 262K context with cheap KV.
Specs
| Architecture | Qwen3_5ForConditionalGeneration (text) |
| Layers | 32 (24 linear + 8 full attention) |
| Quantization | 4-bit affine, group 64 (from BF16 source) |
| Size | 5.6 GB |
| Context | 262,144 tokens |
| Vocab | 248,320 |
Benchmarks (from the distilled base model card)
- GSM8K: 94.50% (teacher DeepSeek-V4-Pro: 92.60%)
- MMLU-Pro (Math/Physics/Chemistry): 90.53% avg — best among 9B baselines tested
- 316.8 correct answers per 1M reasoning tokens; 36.1% fewer tokens per correct
answer vs Qwen3.5-9B
Not a coding specialist — no coding data in the distillation set.
Speed (Mac Studio M4 Max, 128 GB, NovaMLX server)
| Memory resident | ~4.8 GB wired |
| Time to first token | ~1.9 s |
| Decode | ~32 tok/s |
Usage
Any MLX runtime: NovaMLX, mlx-lm, LM Studio.
mlx_lm.generate --model digiflow/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-MLX-4bit --prompt "..."