Liquid AI has posted an 8‑bit export of its d1‑omni‑600M decision model on Hugging Face. The version, called d1‑omni‑600M‑MLX‑8bit, runs on Apple Silicon using the MLX‑vlm library and is only 0.73 GB, about 60 % of the original size. It can answer several named, typed questions about a text, image or audio input in one forward pass, returning no generated tokens. The repo includes a pip install command and example code for text, image and audio queries.
Why it matters
Developers can now run multi‑modal decision models locally on Macs without needing a GPU, enabling faster edge applications like automated support triage.
