Skip to content
H3H3 Max Studio
Sign in

Open weights

MiniMax H3 has open weights — can you self-host?

MiniMax released the H3 weights under the MiniMax H3 Community License. Not everything is open, though. Here's the reality check for anyone tempted to skip the API.

Try the hosted generator →

What's actually available

  • MiniMax H3 (open weights): a 33-billion-parameter model published on Hugging Face (MiniMaxAI/MiniMax-H3) under the MiniMax H3 Community License. It generates native 480p/768p video with native stereo audio, at 4–15 seconds and 24 FPS per the model card — fal's hosted endpoints enforce 5–15 seconds.
  • H3-Regenerate-2K: the module that regenerates a 768p result at 2K is not open-sourced — MiniMax offers it as a hosted API and says it will release it when ready.
  • H3-Context-IR: MiniMax's preprocessing/orchestration system for multimodal inputs is also hosted-only, not part of the open release.
  • H3 Max / H3 Max Turbo: fal's post-trained variants — no separate weights release listed, API only.

Open weights ≠ open source

"Open weights" means you can download and run the checkpoint. It does not mean a permissive license: the MiniMax H3 Community License has territorial and commercial restrictions. Review the full license before self-hosting or using it commercially.

Read the MiniMax H3 Community License →

So: self-host or use the API?

The H3 checkpoint is a 33B-parameter BF16 model. MiniMax's own serving example uses multiple GPUs (four in its SGLang guide) for a single checkpoint, so self-hosting is a real hardware commitment. For most builders, the API is the faster path: fal serves H3 Max Turbo with no infrastructure to provision, and this site generates through that endpoint.

  • Self-host H3 when you need full control of the weights and have the GPU fleet for it.
  • Use the API (H3 Max Turbo) for interactive tools, batch runs, and anything where render time governs the work.

Our approach: generate through fal today, publish transparency pages like this one, and keep the self-hosting facts on file for the day your volume justifies a GPU node.

Read next

Model, capabilities & limits

The model this site actually runs, what is only available upstream, and the sources we verified — so you can check them yourself.

Model used here
MiniMax H3 Max Turbo — fal's post-trained variant of MiniMax H3, served through fal.ai's minimax/h3-max-turbo text-to-video and image-to-video endpoints.
Naming
MiniMax H3 is the open-weight base video model. H3 Max is fal's post-trained variant. H3 Max Turbo is fal's higher-throughput variant. All three are video models — unrelated to any language model.
Resolution
480P and 768P, both native. The base H3 also offers 2K/4K upscales of a 768P render; H3 Max and H3 Max Turbo do not.
Duration
5–15 seconds at 24 FPS (fal's H3 Max Turbo endpoints; the open-weight H3 model card lists 4–15 seconds).
Aspect ratios here
16:9, 9:16, 1:1, 4:3, and 3:4. The upstream schema also lists 21:9, which this site does not expose.
Audio
Native audio is generated together with the video, so sound and picture stay in sync.
Image input
Paste a public image URL (used as the first frame). The upstream endpoint also accepts an optional last frame via end_image_url — not exposed in this site's UI.
Reference-to-video
Available upstream on H3 and H3 Max (H3 Max's reference endpoint arrived after its launch). This site does not offer a reference-to-video mode.
Generation speed
Vendor-reported only. fal's API returns a timings.inference field (backend denoising time); end-to-end wait adds queueing, prompt expansion, and delivery. This site runs no latency benchmark of its own.
Weights
H3 has open weights (33B parameters, MiniMax H3 Community License). H3 Max and H3 Max Turbo list no separate weights release — API only.

Sources

Last verified: September 6, 2026 · Latency figures above are vendor-reported; we publish no site-side measurements.