# MiniMax FastH3 API

> MiniMax FastH3 as a hosted realtime API. Video on main_video and generated audio downmixed to mono on main_audio, aligned per window. $0.0069 per billable second.

Source: https://windflow.online/models/fast-h3 · Realtime video generation · Windflow Labs

## What it is

MiniMax FastH3 is the 35-billion-parameter MiniMax H3 text-to-video-and-audio model, distilled to four DiT calls and sparse attention with the open-source FastVideo stack from Hao AI Lab. On Windflow it streams video with generated audio window by window from a text prompt that carries visual, dialogue, soundscape and music direction.

## Specifications

- **Architecture:** 35B MiniMax H3 text-to-video-and-audio model, DMD2-distilled to four DiT calls with 90% sparse VSA-H3 attention
- **Resolution:** 960 × 544
- **Playout:** 24
- **Chunk:** 345 frames per window (14.38 s at 24 fps; published as the 15 s profile)
- **Control latency:** A whole window. The pipeline materialises the clip before delivery, so time-to-first-frame equals one generation.
- **Inputs:** A text prompt with visual, dialogue, soundscape, and music direction
- **Outputs:** Video on main_video and generated audio downmixed to mono on main_audio, aligned per window
- **Delivery:** WebRTC
- **Session protocol:** sdk.realtime.v2
- **Price:** $0.0069 per billable second

## Pricing

$0.0069 per billable session-second ($25.00/hr of billable time). Billing starts at the session's first generation_started event and stops when the session is released or ends. Queueing, worker startup, time before generation starts and the documented reconnect grace are not billed. After generation starts, paused or unwatched time remains billable; release the owner session to stop the meter.

## Use cases

- Generated video with synchronised dialogue, ambience and music from one prompt.
- Continuous AI channels and episodic streams built window by window.
- Short-form portrait or landscape clips for social and marketing tests.
- Evaluating MiniMax H3 behind a session API before committing to it.

## Where it comes from

- [hao-ai-lab/FastVideo on GitHub](https://github.com/hao-ai-lab/FastVideo) — Apache-2.0 (framework)

## FAQ

### Which licence covers FastH3?

The MiniMax H3 Community License governs the model, including its territory and AI-generation disclosure terms. Read it before building a product on FastH3.

### Can I steer FastH3 mid-window?

No. Each command lands at the next window boundary; prompt changes wait for the window in flight. Soft continuations carry the previous window's final frame forward.

### How much does MiniMax FastH3 cost on Windflow?

MiniMax FastH3 costs $0.0069 per billable session-second ($25.00/hr of billable time). Billing starts at the session's first generation_started event and stops when the session is released or ends. Queueing, worker startup, time before generation starts and the documented reconnect grace are not billed. After generation starts, paused or unwatched time remains billable; release the owner session to stop the meter. There is no subscription.

### How do I get access to MiniMax FastH3?

Windflow is in private beta. Request access from the sign-up page, or ask for a live demo of MiniMax FastH3. Approved accounts start with $5 and four concurrent sessions.

### Is MiniMax FastH3 always running?

No. Model capacity is brought up on demand rather than kept standing. Live readiness for each model is published at api.windflow.online/v1/status, and a live demo can be arranged on request.

## Related guides

- [Choosing a realtime video model: a decision guide for 12 hosted models](https://windflow.online/guides/choosing-a-realtime-video-model)
- [How per-second pricing works for realtime video, with worked examples](https://windflow.online/guides/how-per-second-pricing-works)
- [How to measure realtime video API latency and responsiveness](https://windflow.online/guides/measure-realtime-video-latency)
- [Realtime video API: generate, edit and control live AI video](https://windflow.online/guides/realtime-video-api)
- [How realtime video models work: autoregressive diffusion, chunks and the latency budget](https://windflow.online/guides/how-realtime-video-models-work)