Changelog¶

There are no tagged releases yet. This page is built from the commit history on main, newest first. Dates are commit dates.

2026-10-09¶

README rewrite: plain summary, demo at the top, a "What I built" section listing each file, and corrected setup steps, including fetching reference/ and the optional extras for video, resampling and the parity tests. (bc0e627, merged in 9702e3d)

README: other MLX routes (mlx-vlm, mlx-audio, NVIDIA's own release) and how to report a parity problem. (fedd022)

Added the write-up draft in article/ (Markdown, HTML, build script, poster and demo video). (0c6755f)

README: running the abliterated builds on llama.cpp via GGUF, with the conversion patch, credited to @wacomctl672. (80902bf)

README: links to abliterated (refusal-removed) MLX builds in 4-bit, 8-bit and bf16. Vision and audio towers are left intact. Refusal was removed from blocks 16 and 31 of the language stack. (a16edf0)

README: documented the Metal wired-memory ceiling on 32 GB Macs and the iogpu.wired_limit_mb fix, and moved the warning next to the peak memory numbers. (859d727, 2b45048)

First release of the runtime: src/vision.py, src/audio.py, src/processing.py, src/omni.py, scripts/setup_weights.py and the parity suite in tests/. The commit message reports 23/23 parity tests, vision graph-exact on MLX's CPU stream (cosine 1.00000000), audio cosine 0.99999130, and exact processor token ids. (c4188bd)
Added the demo video, poster and preview GIF in docs/. (8250dd4)
README demo link points to the write-up page, since GitHub shows no player for an mp4 in a repo. (c5df228)