Qwen3.6-27B uncensored heretic v2 — oQ4 (MTP preserved)
Mixed-precision quant of llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, produced with oQ (oMLX v0.5.4). MTP head preserved. Standard MLX safetensors — compatible with oMLX, mlx-lm, LM Studio, and any MLX-capable app.
What is oQ?
Unlike uniform 4-bit quantization, oQ is a data-driven mixed-precision quantizer that calibrates per-layer sensitivity and allocates bits where they matter most. Critical layers (embeddings, LM head, the most sensitive transformer layers) are automatically promoted to 8-bit, while less sensitive layers stay at 4-bit. Typical result: ~4.6 bits-per-weight.
Benchmarked on Qwen3.5-35B-A3B (oMLX project):
| Benchmark | mlx-lm 4-bit | oQ4 |
|---|---|---|
| MMLU (300) | 79.7% | 83.3% |
| TruthfulQA (300) | 87.7% | 88.0% |
| HumanEval (full) | 87.2% | 85.4% |
| MBPP (300) | 71.7% | 74.3% |
Performance (oMLX on M4 10-core)
| Context | PP tok/s | TG tok/s | Peak Mem |
|---|---|---|---|
| 1k | 62.1 | 12.4 | 16.7 GB |
| 4k | 60.8 | 11.5 | 18.2 GB |
| Batch | TG tok/s | Speedup |
|---|---|---|
| 1× | 12.4 | 1.00× |
| 2× | 11.8 | 0.95× |
| 4× | 42.2 | 3.40× |
💡 MTP preserved — compared to the non-MTP quant of the abliterated variant (6.5 tok/s), this model is nearly 2× faster on token generation.
Why this quant
The original BF16 weights require ~55 GB. This oQ4 quant runs in ~16–18 GB on Apple Silicon while keeping the full MTP stack intact.
Quick start
# oMLX
omlx serve --model underlotus/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-oQ4-mtp
# mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("underlotus/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-oQ4-mtp")
response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)
print(response)
Original model
- Base: Qwen/Qwen3.6-27B
- Uncensored: Heretic v2 MPOA pipeline, 94% fewer refusals, KL divergence 0.0021
- MTP: Multi-token prediction head preserved
License
Apache 2.0, inherited from base model.