MiniMax-H3-FP8-Paiton-RDNA4 / MODIFICATIONS.md
Eliovp's picture
Add files using upload-large-folder tool
9cb3b0c verified
|
Raw History Blame Contribute Delete
1.47 kB

Modified files

Every weight file here is modified from MiniMaxAI/MiniMax-H3 (Ref2VA checkpoint) by Paiton, November 2026.

File Source Change
weights/turbo4/dit.bin, dit.idx Ref2VA transformer + Turbo4 LoRA minimax_h3_ref2v_turbo_4step_v0.1 (lightx2v/Minimax-h3-Turbo, MiniMax H3 Community License) LoRA merged into the BF16 weights; the 200 main projections rounded to FP8 e4m3 with one scale per output channel; AdaLN modulation replaced by tables computed for the fixed Turbo4 timesteps; repacked for the Paiton runner
weights/turbo8/dit.bin, dit.idx Ref2VA transformer + Turbo8 LoRA minimax_h3_ref2v_turbo_8step_v1.0_768p as Turbo4, for the Turbo8 timesteps
weights/*/vae_latents.bin video VAE config latent mean and standard deviation, float32
dec8.bin video VAE decoder INT8 weights with block scales, repacked
enc.bin video VAE encoder BF16, repacked
audio.bin audio VAE FP32, repacked
model/tokenizer/tokenizer.json Ref2VA tokenizer unchanged copy

The text encoder files are in EliovpAI/MiniMax-H3-Image-FP8-Paiton-RDNA4 and are described there.