MiniMax-H3 Turbo LoRA (Pruned-Base Compatible, File-Fixed)
Fixed copy of Abiray/MiniMax-H3-Turbo-Lora-Pruned-ComfyUI
(4-step turbo LoRA, original trainer larryvrh/MiniMax-H3-Turbo-Lora).
Why this repo exists
Every .safetensors in the source repo has 64 extra trailing bytes after the last tensor
(declared data ends at byte 620,285,592; files are 620,285,656 bytes), which makes strict safetensors
loaders fail with:
SafetensorError: Error while deserializing header: incomplete metadata, file not fully covered
This file is byte-identical to the source file with the 64 trailing bytes removed. All 416 tensors
(attn.qkv_proj / attn.out_proj / mlp.fc1 / mlp.fc2 per block × 50 blocks + token_refiner,
no adaln targets — designed for pruned / curve-form MiniMax-H3 checkpoints) load and match a
pruned FP8 base (e.g. Comfy-Org/MiniMax-H3 *_pruned_fp8_scaled).
Usage (ComfyUI)
Load LoRA node → minimax_h3_turbo_4step_ckpt600_ema_V4_fixed.safetensors, strength 0.8–1.8,
Steps 8–12, sampler res_multistep. (Settings from the author's workflow.)