library_name: mlx
license: apache-2.0
base_model:
- huihui-ai/Huihui-Qwen3.8-27B-abliterated
- harness-hub/Qwen3.8-27B-pi-adapted
pipeline_tag: image-text-to-text
tags: - mlx
- omlx
- oq
- oQ6e
- quantized
- imatrix
- qwen3_5
- mtp
- pi
- agents
- abliterated
- uncensored
Qwen3.8-27B-pi-abliterated-oQ6e-fp16-mtp
huihui-ai/Huihui-Qwen3.8-27B-abliterated with the harness-hub/Qwen3.8-27B-pi-adapted LoRA adapter merged in, quantized for Apple Silicon with oMLX oQ6e. The adapter was trained with reinforcement learning for the pi coding agent; see its card for the method and the benchmark results, which were measured on the unquantized, non-abliterated model.
The non-abliterated sibling is grokked-it/Qwen3.8-27B-pi-oQ6e-fp16-mtp.
The checkpoint holds the language model, the vision encoder and the multi-token prediction (MTP) head.
How it was made
- Merge. The adapter (LoRA, rank 32, alpha 64, 256 linear layers of the language model) was merged into the abliterated BF16 weights:
W' = W + 2 · B · A, computed in float32 and stored back in BF16, one shard at a time. The adapter was trained on the original Qwen weights, not on the abliterated ones. The vision encoder and the MTP head are the abliterated repository's, which its card says are unmodified from the base model. - Quantize. oMLX 0.8.0, oQ level 6, enhanced (
e): mixed precision by layer sensitivity, importance matrix on theoqe_code_multilingualcalibration set (128 samples of 512 tokens, adaptive), code refinement. Scales, biases and unquantized tensors in FP16, MTP head preserved.
| Base bits | 6-bit affine, group size 64 |
| Tensors raised to 8 bits | 44 |
| Size | 24.8 GB |
| Estimated effective bits per weight | 6.87 (oMLX estimate before the run) |
Tokenizer and chat template come from the adapter repository.
Usage
Load it in oMLX, or with any MLX loader that supports the Qwen3.5 architecture.
The adapter's authors recommend thinking off ("chat_template_kwargs": {"enable_thinking": false}) and report their results at temperature 0.8, top-p 1.0. Tool calls use the base model's format.
This model has its refusal behaviour removed and will comply with requests the base model declines. You are responsible for how you use it.
Limits
- No benchmark was run on this checkpoint. It passed only a smoke test in oMLX: one code completion, one tool call, and two prompts that the non-abliterated sibling refused and this model answered.
- Two prompts do not measure how complete the abliteration is after the merge.
- The MTP head was trained against the original backbone, not the adapted or abliterated one.
- The merge rounds to BF16 before quantization, as the adapter card's own merge recipe does.
License
Apache-2.0, following the base model, the abliterated weights and the adapter. See LICENSE.