MiniMax H3: Open‑Source Next‑Gen General Video Model Matching Seedance 2.0

MiniMax has open‑sourced its latest general‑purpose video model, H3, which achieves performance comparable to Seedance 2.0 and ranks among the top three worldwide for text‑to‑video and image‑to‑video generation.

H3 is a full‑modal model that can understand and generate text, images, video and audio, outputting native stereo 2 K video up to 15 seconds, with multiple aspect‑ratio options and support for 11 languages.

One Model, All Modalities

Previously, content‑generation pipelines were fragmented into separate models for text‑to‑image, audio synthesis, video‑to‑video motion transfer, etc., limiting flexibility and generalisation. H3’s design unifies these tasks, allowing a single prompt that may include text, reference images, video clips and audio to produce a coherent video‑audio output.

Three Dedicated Modules

H3 consists of:

H3‑Context‑IR : parses and aligns multimodal inputs, performs instruction parsing, cross‑modal association, temporal reasoning and produces a structured intermediate representation; currently provided via API.

H3‑Base : the core generation engine that encodes text, visual and audio inputs (via H3‑Encoder, H3‑VisualVAE, H3‑AudioVAE), packs them into a multimodal token sequence, and feeds them to an Omni‑Transformer which predicts video and audio latents; supports sparse attention for long sequences.

H3‑Regenerate‑2K : upscales the 768 p output to 2 K by feeding the low‑resolution result together with the original context back into H3, preserving fine details such as small text.

Running the Model

Two task‑specific checkpoints are released on Hugging Face: H3‑Base‑FL2VA for first/last‑frame generation and H3‑Base‑Ref2VA for full‑modal reference mode (up to 9 images, 3 video clips, 3 audio clips, max 12 files). Each checkpoint includes processor, tokenizer, encoders, transformer and VAEs, and can be used with SGLang, vLLM, Diffusers or ComfyUI.

MiniMax also provides three reproducible 768 p generation examples (T2VA, FL2VA, Ref2VA) and a “Full 2K Workflow” that chains Context‑IR, H3‑Base and Regenerate‑2K to obtain 2 K video.

Commercially, H3 shows strong instruction following, brand‑information rendering, and video‑to‑video motion transfer, making it suitable for advertising, e‑commerce, product design, UI/UX and game content.

All model weights, documentation and prompt‑writing guides are available on Hugging Face, and the API reference is published on MiniMax’s platform.