Qwen3.8-27B Abliterated Released for Apple Silicon via MLX

X postTrevor WoodAnnouncement

Trevor Wood and PocketAI Model Lab have released MLX-optimised checkpoints of Qwen3.8-27B for Apple Silicon with built-in refusal mechanisms removed ('abliterated'), provided across multiple quantisation formats from 2-bit to BF16.

Key points

  1. Variants include experimental 2-bit AWQ (10.28 GiB), 4-bit, 6-bit, 8-bit, and reference BF16 (50.98 GiB).
  2. The 4-bit, 6-bit, 8-bit, and BF16 models scored 0/100 explicit refusals on test benchmarks while passing 12/12 quality tests and 8/8 native tool tests.
  3. The 2-bit AWQ build targets 16 GB RAM Macs but failed native tool-calling checks and showed lower quality scores.
  4. On an M5 Max with 128 GB unified memory, the 4-bit version achieved 33.2 tok/s generation throughput with 21.80 GB peak memory.
  5. PocketAI Model Lab issues a safety notice stating the checkpoints may output harmful or incorrect content more readily than the base model.

Why it matters

MLX quantisation brings 27-billion-parameter multimodal models directly to local Apple Silicon hardware. Abliterated weights provide unrestricted outputs needed for research and synthetic data pipelines, though they shift safety responsibilities entirely to the operator.

Analysis

The abliteration recipe projects out refusal directions across layers 24 to 63 across 80 residual matrices, preserving visual analysis and tool integration in versions rated 4-bit and higher. However, heavily compressed 2-bit implementations break complex features like tool execution and exact needle retrieval.

Original post

Trevor Wood

@trevorwood222 · 15 August 2026

Qwen3.8-27B Abliterated is now available for Apple Silicon with MLX 🔓🍎 Separate 4, 6, 8-bit + BF16 releases. All four scored 0/100 explicit refusals on harmful prompts and benign controls, plus 12/12 quality, 8/8 tools and verified 4K retrieval. t.co/vUgQBjbuAa

2.8k likes50 replies3.2k bookmarks199k views

View on X
More posts in the thread (10)
  1. @AnxKhn I’ll add a 1 & 2 bit in today 🫡
  2. @AnxKhn 1 bit looks a bit difficult, will do 2 bit first 🫡 Your 16gb ram Mac should be able to run it
  3. @MarkusBaue15570 @AnxKhn Can probably get 9-14 tok/s on an M1 at 2 bits That’s pretty useable for an agent
  4. @AnxKhn Uploading 2 bit now. Should work on 16gb macs. Got 4/200 refusals for it
  5. @Peter_gcg @grok Basically will just do it’s best to answer your questions without determining whether or not it should or should not.
  6. @wessorh What do you mean? You can give it your own system prompt
  7. @veilgnostic @NVIDIAAI could I have a DGX spark to make models on please 👍
  8. @es861vv6 The 2 bit can probably run well. Just uploaded it
  9. @Omar_Sultan Seems that LM Studio needs to update their MTPLX version for it to work. I think there’s a pull request open Also doing tests. Seems MTPLX is only better for small context, regular is better for longer context Testing an updated MTPLX version that might do better
  10. @Miss1ngFrog Dang 10 tok/s is slow Is that on the 4bit? How much context? I’m testing a new MTPLX version now. Might be faster. I’ll test it on my M1 Max 32GB to see how it does
Selected replies (5)
  • @AnxKhn @trevorwood222 need unsloth's 1 bit for my 16 GB Mac :(
  • @JJeffrey100 @trevorwood222 i even get 4 bit mlx running on a 24GB MBP (M4 Pro). had to cut context to 8k, so it's not doing code, but is very capable for even vision MUCH better on 32GB and flawless at 64GB. I have my studio running a refactor on a test codebase so will know more in about an hour
  • @Peter_gcg @trevorwood222 @grok whats is an abliterqted version of a model and how is this significant
  • @KalenuikMatthew @wessorh @trevorwood222 @wessorh the prompt does not stop refusals for unsafe or restricted model behavior. Even a system prompt that explicitly instructs it to do anything the user asks, will not result in it doing what the user asks if the alignment training would reject the request. Therefore, they
  • @julianharris @JulianRaxworthy @trevorwood222 Synthetic data is a great example: I have a moderation tool that looks for abusive or disrespectful language, and tools like that by definition will be pushing boundaries of censorship. And any Chinese models have their own communist agenda. I saw a code snippet that actively

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Link:PocketAiHub

    Qwen3.8-27B-Abliterated-MLX: Uncensored Multimodal Model for Apple Silicon

    PocketAiHub has released an unofficial MLX derivative family of Qwen3.8-27B with suppressed refusal behaviour ('abliteration'). The consolidated repository provides quantized builds ranging from experimental 2-bit AWQ to full BF16 optimised for Apple Silicon.

    AI & LLMs· Resource

  • Link:OrcaRouter

    OrcaRouter Releases Abliterated FP8 Build of Qwen3.8-27B

    OrcaRouter has released an abliterated, block-FP8 quantized version of the multimodal model Qwen3.8-27B on Hugging Face. By orthogonalising the internal refusal direction, standard safety guardrails were removed specifically for red-teaming, interpretability, and AI safety research.

    AI & LLMs· Tool

  • Link:Unsloth AI

    Unsloth Releases Dynamic V3.0 GGUF Quants for Qwen3.8-27B

    Unsloth AI has published GGUF quantizations for the Qwen3.8-27B model under an Apache-2.0 licence. Utilizing their new Dynamic V3.0 quantization format, Unsloth claims over 10% higher top-1 accuracy at identical sizes compared to alternative quants. Qwen3.8-27B is a 27-billion parameter dense vision-language model featuring native text, image, and video comprehension alongside granular thinking controls.

    AI & LLMs· Tool

  • Link:Unsloth AI

    Unsloth Releases Qwen3.6-27B-MTP in GGUF Format

    Unsloth AI has published Qwen3.6-27B-MTP-GGUF, providing quantized versions of the first open-weight model in the Qwen3.6 family. The 27-billion parameter multimodal model features Multi-Token Prediction (MTP) for accelerated inference alongside upgrades to agentic coding and tool calling.

    AI & LLMs· Resource

  • Repository:PipeNetwork/kimi-k3-mlx

    kimi-k3-mlx: Apple Silicon Port of the 2.78T Multimodal MoE Kimi-K3

    kimi-k3-mlx is an Apple MLX port of Moonshot AI's native multimodal MoE model Kimi-K3 (2.78T total parameters). Because the model exceeds the memory capacity of any single Apple Silicon machine, the repository includes a custom streaming converter, vision components, and expert-pruning tools.

    337starsPython

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.