Realtime models

MiniMax H3

Open-weight 33B video model (Hailuo 3.0) — 15s clips with native stereo audio; the substrate under this wave.

huggingface.co · listing ≠ endorsement · re-checked 2026-09-03

Params: 33B dense · open base weights (H3-Base)
Output: 4–15s clips; 2K via API only (local ≈768p)
Audio: Native stereo — a property of the model
Weights opened: Aug 3, 2026
Licence note: Regional exclusions incl. US / EU / UK / KR

What is MiniMax H3?

The open-weight 33B video model (Hailuo 3.0) that much of this wave stands on. Only the base weights are open (locally around 768p); the 2K upscale pass and context-reference modules are API-only, and the licence carries regional exclusions. MiniMax's own API is not realtime — but fal's live-tuned build turns H3 into a streaming substrate.

What MiniMax H3 does

33B dense open-weight model (H3-Base) released Aug 3, 2026
Generates 4–15 second clips with native 32kHz stereo audio — audio is in the model, not post-processed
Omni-modal: text, image and video in; video and audio out
2K upscale and context-reference modules are API-only — local open-weight generation is around 768p
Day-zero ecosystem support from Hugging Face, ComfyUI, ModelScope and many chip vendors

Who is MiniMax H3 for?

Self-hosting open-weight video generation on a local GPU
Building video pipelines that need native, controllable audio
Running realtime or near-realtime generation via fal’s H3 Max build

Pros and cons of MiniMax H3

Pros

One of the first flagship open-weight video models with native audio
Runs locally — ComfyUI shipped day-zero support and cites RTX-3060-class GPUs
Day-one adaptation across major open-source tooling and silicon vendors

Cons

Only the base is open; the 2K pass and reference modules are API-only
Licence carries regional exclusions (incl. US, EU, UK, South Korea)
MiniMax’s own API is not realtime — the streaming use case runs through fal’s tuned build