A standalone text-only MLX 4-bit release published by Interim Labs for local text generation on Apple Silicon. This package derives from huihui-ai/Huihui-Qwen3.5-2B-abliterated. Interim Labs packages and documents this derivative; the upstream model and conversion sources are credited below.
Part of the Interim Labs Qwen3.5 series: three sizes, each with an original Qwen3.5 variant and a Huihui abliterated variant, in two formats.
| Model | MLX 4-bit | GGUF Q4_K_M |
|---|---|---|
| Qwen3.5 2B | MLX | GGUF |
| Huihui Qwen3.5 2B abliterated | MLX | GGUF |
| Qwen3.5 4B | MLX | GGUF |
| Huihui Qwen3.5 4B abliterated | MLX | GGUF |
| Qwen3.5 9B | MLX | GGUF |
| Huihui Qwen3.5 9B abliterated | MLX | GGUF |
| Property | Value |
|---|---|
| Parameter class | 2B |
| Format | MLX / Safetensors |
| Quantization | 4-bit affine, group size 64 |
| Weight file | model.safetensors |
| Weight size | 1.06 GB (1,059,394,783 bytes) |
| Model type | qwen3_5_text |
| Configured context | 8,192 tokens |
| Retained text tensors | 694 |
This package contains language-model tensors only and does not provide image input. Text-only derivation removed 297 vision tensors. The packaged configuration sets the context to 8,192 tokens; this is not the upstream model’s full advertised context length.
Download the package with the Hugging Face CLI:
hf download interimlabs/InterimLabs-Huihui-Qwen3.5-2B-abliterated-MLX-4bit
Use an MLX runtime on Apple Silicon that supports the qwen3_5_text architecture and this package’s quantization. Use the included tokenizer and chat template, and keep the total context within 8,192 tokens. Generic Python mlx_lm builds without qwen3_5_text support may fail to load this package. Runtime compatibility across applications has not been comprehensively verified.
b2e291a65f29a9b148981fa5299caea5d35bd4c8.mlx-vlm==0.3.12 and mlx==0.32.3, followed by text-only derivation.b2e291a65f29a9b148981fa5299caea5d35bd4c8.The Huihui upstream identifies Qwen3.5-2B as its base but does not specify the base revision or a complete abliteration recipe.
See provenance.json for source and package metadata.
Weight-file SHA-256:
67a627fad9153d3df31bf8e86449516663ed86db9e53fdfc1334435c05d042c2
An earlier local integration using mlx-swift-lm 3.31.3 loaded and generated with this package. In a small internal set of writing-operation checks, it passed 5 of 7 checks, compared with 4 of 7 for the original Qwen3.5-2B package. Observed errors included an incorrect proofreading suggestion, incomplete JSON, and incorrect dialogue attribution. Both variants answered four ordinary romance and war-fiction prompts without refusing. This small check is not a general benchmark and does not establish an improvement in refusal behavior.
The Huihui variant uses an upstream abliterated checkpoint. This release does not establish a measured refusal-rate improvement, and abliteration should not be taken as a guarantee of answer quality.
Distributed under Apache-2.0. Credit belongs to the Qwen team, Huihui, and the conversion authors linked above. Interim Labs publishes this package and its documentation. See the linked upstream cards for the original model details and limitations.
huihui-ai/Huihui-Qwen3.5-2B-abliterated