Meridian-H3-8bit-MLX / BASE-MODIFICATIONS.md
Avdpro's picture
Meridian MLX 8-bit composite; original territorial license applies
78df34d verified
|
Raw History Blame Contribute Delete
1.63 kB
# Modifications to MiniMax H3
These files are MODIFIED versions of the MiniMax H3 Works, redistributed under
the MiniMax H3 Community License Agreement (see LICENSE and NOTICE).
Modified by: mlx-serve (https://github.com/ddalcu/mlx-serve)
What changed:
* `transformer.safetensors` and `text_encoder.safetensors` are QUANTIZED from
the original bfloat16 releases to MLX affine 8-bit, group size 64. Gathered
embedding tables and the checkpoint's fp32 islands (patch projections, output
heads, time embedder, rope inverse frequencies) are left dense.
* `video_vae.safetensors` and `audio_vae.safetensors` are byte-for-byte copies
of the originals, unmodified.
* The tokenizer files are byte-for-byte copies from `MiniMaxAI/MiniMax-H3`
(`FL2VA/processor/`), relocated into this directory so the model is
self-contained.
* `config.json` is new, written by mlx-serve's converter to describe the
layout above. It is not from the original release.
* `turbo_lora.safetensors` is NOT from MiniMax and is not a modification of
their Works. It is an unmodified, renamed copy of
`minimax_h3_turbo_4step_ema_ckpt850.safetensors` from
https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora, Copyright (c)
larryvrh, licensed under the Apache License, Version 2.0. Being a LoRA
trained on MiniMax H3, it is also a derivative of the Works and remains
subject to the MiniMax H3 Community License Agreement.
No MiniMax weights were retrained, distilled, pruned or otherwise altered
beyond the numeric quantization described above. The Turbo adapter is a
separate file that is added at runtime; it does not modify them.