High‑quality video generation benefits from systems that actually understand mixed multimodal context rather than treating each modality separately. MiniMax H3 pursues that direction by combining a hosted Context‑IR preprocessing stage with an open H3‑Base generator and an API‑hosted in‑context 2K regeneration flow, letting you author temporally coherent video with synchronized stereo audio from rich text/image/video/audio prompts.
Key Capabilities
- Multimodal inputs and modes: text‑to‑video, first/last‑frame to video (FL2VA), and reference‑to‑video/audio (Ref2VA) that accepts images (≤9), short clips (≤3, total ≤15s) and accompanying audio.
- Output specs: 4–15 second outputs, flexible aspect ratios (including 16:9/9:16), 24 FPS, 32 kHz stereo audio; H3‑Base produces 768p results and an in‑context H3‑Regenerate‑2K module upsamples to 2K while reusing original context.
- Architecture highlights: single‑stream H3‑Omni‑Transformer (~33B params) with modality‑agnostic blocks, AdaLN modality branches, MM‑RoPE positional encoding; H3‑Encoder uses Qwen3‑VL‑32B hidden states; H3‑VisualVAE is a temporally causal f16t4d24 video VAE; H3‑AudioVAE encodes stereo 32 kHz audio at 40 Hz latent rate.
- Distribution and deployment: two open checkpoints (MiniMax‑H3 Base FL2VA and Ref2VA) released for local inference (BF16), with diffusers, SGLang and vLLM serving recipes; Context‑IR and the 2K regeneration service are provided as hosted APIs for full‑quality workflows.
Who it's for and trade‑offs
Great fit if you need a single model pipeline to generate temporally coherent video with synchronized stereo audio from complex multimodal prompts or reference material, and you can provision multiple GPUs for inference. It is also useful for teams wanting a reproducible local 768p generator and an official API path to validate 2K outputs. Look elsewhere if you require a lightweight on‑device text‑only T2V model, deterministic low‑cost frame synthesis for long clips, or fully open-source end‑to‑end 2K regeneration (H3‑Context‑IR and H3‑Regenerate‑2K remain hosted/not yet open‑sourced). Expect nontrivial GPU and engineering costs to deploy at scale (examples use multi‑GPU SGLang setups). Also note the model is distributed under the MiniMax H3 Community License, which imposes use restrictions beyond a permissive OSS license.
