GitHub radar
Qwen 27B on Any MacBook: 8.6 GB on Disk
Prism ML compressed Alibaba's Qwen3.8-27B model down to 8.6 GB using ternary weights, making it run natively on MacBooks with Apple Silicon.
Prism ML compressed the Qwen3.8-27B model from Alibaba using ternary (2-bit) weights. The resulting file is 8.60 GB on disk and runs natively on Apple Silicon. The authors report 98.2% of the original quality is retained. The model understands images alongside text, writes and explains code, handles long conversations, and supports tool use for agentic tasks. On an Apple M5 Max it outputs around 47 tokens per second. This format works only on Mac computers with M-series chips.
Why a vibe-coder should care
The authors describe it as deployable on everyday laptops at 8.60 GB on disk. Most models at this capability level require 32 GB of memory or a dedicated GPU. For tasks like answering questions, writing code, reviewing images, or running as a local agent, this is one of the few 27-billion-parameter models that can fit on a 16 GB MacBook. It works offline, without a paid API subscription.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the compressed Qwen model by Alibaba for my Mac: https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit - follow the README instructions, install the dependencies, and run a test conversation.
A Mac with Apple Silicon and at least 16 GB of unified memory. The 2-bit model takes 8.60 GB on disk. It runs on Apple Silicon only.
Open on Hugging Face▌ More finds