H3 Video Model Backups
Personal backup mirror of two large files freed up from a local disk during H3 video-generation experiments (MiniMax H3 / Qwen3-VL-32B pipeline work).
minimax_h3_fl2v_turbo_4step_v1.0_merged_cache/
A LoRA-merge cache: the MiniMax H3 FL2VA DiT checkpoint with the
minimax_h3_fl2v_turbo_4step_v1.0_768p LoRA pre-merged and re-quantized
(int8 + scale, per-module state_dict() files). This is a derived
artifact — it can be regenerated from the base DiT checkpoint + LoRA file
in a few minutes; it's kept here purely as a convenience backup, not as
original data.
qwen3vl_32b_h3_ultra_uncensored_heretic_generation_tail_50_63_int8_convrot.safetensors
The "generation tail" (transformer layers 50–63 + final norm + LM head) that
reconnects MiniMax H3's truncated Qwen3-VL-32B text encoder into a full
text/vision-language generator. This is a third-party file, not original
work — originally published by ethanfel for the
ComfyUI-H3-Qwen3VL-TextGen
node pack, at
ethanfel/Qwen3-VL-32B-H3-ComfyUI-Generation-Tails.
Mirrored here for personal convenience; see the original repo for the
authoritative source, license, and updates.