H3-Tiny-17B-Pruned-Ref2VA

Part of H3-Tiny: structured pruning of the MiniMax H3 Ref2VA transformer (33B → 17B → 14B) for joint video + audio generation from text and reference images. Project page · Code · All checkpoints

Parameters 17.07B (transformer)
Workflow Ref2VA (reference images to video + audio)
Made by MiniMax H3 Ref2VA (33.12B) with AdaLN rank 8 and 8 of its 50 transformer blocks removed. No training.
Sampling steps 50

This is the training-free checkpoint: it works straight out of pruning.

This repository holds the transformer only (diffusers layout: config.json + sharded safetensors). The text encoder, VAEs and processors come from the official MiniMax H3 checkpoint in diffusers format.

Usage

git clone https://github.com/fuyi25/H3-Tiny && cd H3-Tiny && pip install -e .
hf download H3-Tiny/H3-Tiny-17B-Pruned-Ref2VA --local-dir ckpt/H3-Tiny-17B-Pruned-Ref2VA
python scripts/infer.py --weights ckpt/H3-Tiny-17B-Pruned-Ref2VA --base-model /models/MiniMax-H3 \
    --prompt "<Picture 1> ..." --references ref1.png --num-frames 124 --out outputs/
from h3tiny.pipeline import load_pipeline
pipe = load_pipeline("ckpt/H3-Tiny-17B-Pruned-Ref2VA", "/models/MiniMax-H3", device="cuda")

See the GitHub README for prompts, offloading, attention backends and the pruning scripts.

License

These weights are Model Derivatives of MiniMax H3 and are distributed under the MiniMax H3 Community License Agreement; see NOTICE. Use of the weights must follow that agreement, including its territorial scope and acceptable-use policy.

H3-Tiny is an independent project and is not affiliated with or endorsed by MiniMax.

Downloads last month
252
Safetensors
Model size
17B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including H3-Tiny/H3-Tiny-17B-Pruned-Ref2VA