Video Generated FasterThan You Can Watch It
MiniMax H3 Max is fal's post-trained build of MiniMax H3 — retrained for sharper prompt adherence and rebuilt for speed. It returns a 5-second 768p clip with synchronised audio in under 3 seconds, roughly 35× the throughput of the stock H3 endpoint.
Why H3 Max Changes the Loop
When a generation returns before you have finished reading your own prompt, video stops being a batch job you queue and becomes something you iterate on live.
A 5-second clip lands in under 3 seconds — the video finishes generating before it finishes playing. That is roughly 35× the throughput of the official H3 endpoint.
In blind human preference evaluations against a field of leading video models, H3 Max ranked first on overall quality, prompt understanding and aesthetics — not just on speed.
fal Research started from H3's open weights and post-trained on substantial new data, with the focus squarely on doing what the prompt actually says and looking good doing it.
H3 Max inherits H3's native audio generation — each clip returns with synchronised sound already attached, so there is no separate audio pass and nothing to conform in an editor.
At $0.05/sec for 480p and $0.08/sec for 768p, a discarded take costs cents. The economics favour generating ten options and keeping one.
Drive a shot from a written description or animate a still you already have. Both modes run on the same fast path, with prompt expansion available when you want the model to fill in detail.
H3 Max vs MiniMax H3
These are two different models with different jobs. H3 Max trades resolution and multimodal conditioning for enormous speed; stock H3 keeps 2K and the full omni-modal input surface.
| MiniMax H3 Max | MiniMax H3 | |
|---|---|---|
| Built by | fal Research (post-trained) | MiniMax |
| Released | 27 August 2026 | 31 July 2026 |
| Max resolution | 768p (480p / 768p) | Up to 2K |
| Speed | 5s clip in under 3s | Standard endpoint latency |
| Relative throughput | ~35× stock H3 | Baseline |
| Input modalities | Text, image | Text, image, video, audio |
| Audio | Synchronised, native | 32 kHz native stereo, 11 languages |
| Clip duration | 5–15 seconds | 4–15 seconds at 24 fps |
| Aspect ratios | 16:9 | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 |
| Best for | Fast iteration, volume, live loops | Finishing, 2K delivery, reference-driven work |
H3 Max does not currently support 2K output or the complex intermediate-frame and multimodal reference inputs available on stock H3. Pick the model that matches the stage you are at.
What a Finished Minute Actually Costs
H3 Max at 480p lands near $3 per finished minute — where mainstream rivals sit above $20.
per second — about $3.00 per minute. Launch promo runs at half this.
per second — about $4.80 per minute at full H3 Max resolution.
per second — about $7.80 per minute when you need true 2K delivery.
| Model | Resolution | Price / second | Price / minute |
|---|---|---|---|
| MiniMax H3 Max | 480p | $0.05 | ~$3.00 |
| MiniMax H3 Max | 768p | $0.08 | ~$4.80 |
| MiniMax H3 | 2K | $0.13 | ~$7.80 |
| Veo 3.1 (standard) | — | $0.40 | ~$24.00 |
| Kling 3.0 | 1080p | — | ~$20.16 |
| Seedance 2.0 | 1080p | — | ~$22.45 |
List API prices as published by each vendor, August 2026. fal is running launch pricing at 50% off H3 Max. Platform pricing on Pollo AI and other resellers differs — free credits are available to start.
How to Work With H3 Max
Explore fast and cheap, then finish where the resolution matters.
Write the shot, or upload an image to animate. H3 Max takes text and image input — the heavier video and audio reference conditioning lives on stock H3.
Because a take returns in under three seconds, you can rewrite and regenerate in a tight loop rather than queueing a batch and coming back after coffee.
When a concept is locked and you need 2K, vertical framing or reference-driven control, take the winning prompt over to MiniMax H3 and render the delivery master there.
Where Speed Actually Wins
Sub-3-second generation does not just make old workflows quicker — it makes a few new ones possible.
Generate fifty ad variants in the time a conventional model renders two, then put the whole set into testing instead of betting the budget on a single concept.
Faster-than-realtime generation puts video inside the request cycle. Build features where a user types and watches, rather than submits and waits for an email.
Find the shot on H3 Max for pennies while the room watches, then re-render the approved prompt at 2K on stock H3. The exploration cost stops mattering.
Animate an entire product library rather than the hero SKUs. At three dollars a finished minute, covering the long tail finally pencils out.
Keep a daily posting cadence fed without a render queue. Clips arrive with audio attached and are ready to cut the moment they land.
Pitch with moving images instead of mood boards. Generate the reference during the meeting rather than promising it for next week.
Frequently Asked Questions
Straight answers about H3 Max, who built it, and how it relates to MiniMax H3.
Ready to Generateat the Speed of Thought?
H3 Max finds the shot in seconds. MiniMax H3 delivers it in 2K with full omni-modal control. Start with the H3 family on Pollo AI — free credits, no credit card.