# H3 Max

> Source: https://aiwiki.ai/wiki/h3_max
> Updated: 2026-08-28
> Fact-checked: 2026-08-28
> Categories: AI Models, Video Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "H3 Max." aiwiki.ai, 28 Aug 2026. https://aiwiki.ai/wiki/h3_max
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**H3 Max** is a hosted video-generation model developed by [fal.ai](https://aiwiki.ai/wiki/fal_ai) as a post-trained derivative of [MiniMax H3](https://aiwiki.ai/wiki/minimax_h3). fal announced the model in August 2026 and made it available through its playground and API. The release combines changes to the base model with a serving stack optimized by fal's inference team.[1][2]

H3 Max is distinct from MiniMax H3. The base model was developed by [MiniMax](https://aiwiki.ai/wiki/minimax), supports 2K output and additional reference and editing workflows, and has publicly released weights subject to its community license. H3 Max instead runs at up to 768p through fal's hosted endpoints. fal had not released H3 Max weights, a model card, or a downloadable checkpoint as of August 28, 2026.[1][2][7]

## Overview

| Property | Detail |
| --- | --- |
| Developer | fal Research |
| Base model | MiniMax H3 |
| Release | August 2026 |
| Access | fal playground and API |
| Launch endpoints | Text-to-video and image-to-video |
| Resolution | 480p or 768p; 768p default |
| Duration | 5-15 seconds |
| Frame rate | 24 fps, according to fal |
| Audio | Synchronized audio generated with video |
| Weight availability | No public H3 Max weights announced as of August 28, 2026 |

fal describes H3 Max as the product of additional post-training on the open-weight MiniMax H3 base. The company says it introduced new data and used preference studies and reinforcement-learning tasks to target prompt adherence and visual quality. It also developed the model alongside a custom inference system rather than adapting the completed checkpoint to a general serving stack afterward. fal has not disclosed H3 Max's parameter count, post-training dataset, training compute, detailed reinforcement-learning method, or architecture changes.[1][2]

The distinction between model and service is important. MiniMax H3's public weights do not establish that H3 Max is [open weight](https://aiwiki.ai/wiki/open_weights). H3 Max is a fal-hosted derivative, and the available documentation describes API access rather than local deployment.[2][7]

## Capabilities and API

H3 Max launched with two endpoint families. `minimax/h3-max/text-to-video` generates video from a prompt. `minimax/h3-max/image-to-video` can use an image as the opening frame and an optional second image as the ending frame. If the image field is omitted from the latter endpoint, fal's schema says that the request is handled as text-to-video.[3][4]

Both endpoints generate 480p or 768p output, with 768p as the default. fal specifies 1344x768 pixels for 16:9 output and a frame rate of 24 fps. Durations range from 5 to 15 seconds. Text-conditioned requests support 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 aspect ratios, while image-conditioned output follows the supplied image's ratio.[2][3][4]

The model produces synchronized audio along with the video, continuing a capability of the MiniMax H3 base. The API accepts a random seed and includes a safety checker that defaults to enabled. Its prompt-expansion setting can be disabled or set to balanced or quality. Balanced is the default; fal says the quality setting can spend about 30 seconds producing a richer prompt before generation.[2][3][4]

Responses contain the generated video as a hosted file, or as base64 data when synchronous mode is selected. They may also include the expanded prompt and an inference-timing object. fal's image endpoint supports first-to-last-frame generation, but the broader reference-to-video and video-editing workflows documented for MiniMax H3 were not available as verified H3 Max endpoints at the August 28 cutoff.[3][4][7]

## Performance claims

fal reports that H3 Max can generate a five-second 768p video in under three seconds of wall time. A sample API response showed approximately 2.53 seconds for the denoising portion of one generation. The company also described this as roughly 35 times the throughput of the official MiniMax H3 endpoint and an average of 15 times the speed of models it considered comparable in quality.[1][2]

Those figures are vendor measurements. fal did not publish a reproducible latency protocol covering hardware, queue time, prompt set, repeated trials, or tail latency. The timing field measures backend denoising and is not necessarily total application latency, which can also include prompt expansion, queueing, upload, and download time.[2][3]

fal also ran an internal human-preference evaluation against twelve video models. Evaluators compared overall preference, prompt understanding, and aesthetics, with results aggregated through Bayesian Elo ratings and 95% confidence intervals. fal reported that H3 Max ranked first in all three dimensions and won most head-to-head comparisons. These are first-party results; the company did not publish the complete prompt set, vote data, or enough procedural detail for independent reproduction.[1]

## Independent leaderboard results

[Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) evaluates [AI video generation](https://aiwiki.ai/wiki/ai_video_generation) systems through blind, crowdsourced pairwise preferences. Its rankings are dynamic snapshots rather than permanent model properties.[5][6]

In a live snapshot retrieved on August 28, 2026, at 05:37:52 UTC, H3 Max ranked first on Artificial Analysis's image-to-video leaderboard with audio. In the separately scored text-to-video leaderboard with audio, it ranked third, behind Wan 3.0 and Gemini Omni Flash. These positions describe two specific leaderboard modes at that timestamp and can change as votes accumulate.[5][6]

This snapshot does not support an unqualified claim that H3 Max was the current number-one text-to-video model. It does support first place on the with-audio image-to-video board at that date. Artificial Analysis normalizes price to a minute of 1080p-equivalent output at default settings, so its displayed $2.40-per-minute figure is a comparison convention rather than H3 Max's native output resolution.[5][6]

Arena results measure preference among generated samples. They do not directly measure factual accuracy, safety, consistency over long sequences, prompt adherence in every domain, or reliability under production traffic.

## Pricing and availability

H3 Max was available through fal's playground and authenticated API as of August 28, 2026. The live text-to-video endpoint listed launch promotional prices of $0.025 per output second at 480p and $0.04 per output second at 768p, with the discount scheduled to end September 1. The same page listed subsequent prices of $0.05 and $0.08 per second, respectively.[2][3]

fal's broader product page contained different launch pricing language, including a $0.06-per-second 768p list rate and a $0.03 promotional rate. Because the endpoint page governs the callable model and supplied an explicit end date, the figures above reflect that page's August 28 listing. Users still need to check the current endpoint before estimating costs.[2][3]

fal also advertised five free sandbox generations per rolling 24-hour period for signed-in users, up to 15 seconds each. That allowance and the launch discount were service terms, not properties of the model, and could change independently of its technical behavior.[2]

## Limitations

Public documentation leaves major technical details undisclosed. It does not identify H3 Max's exact architecture changes, post-training corpus, compute budget, full reinforcement-learning procedure, or checkpoint license. The model is limited to 768p on its documented endpoints, compared with 2K through the base MiniMax H3 service. Its speed comparisons and internal quality study come from fal, the developer and hosting provider.[1][2][7]

The public leaderboard evidence is narrower than fal's marketing claims. H3 Max led the image-to-video-with-audio mode in the cited snapshot but ranked third in text-to-video with audio. Scores, prices, and rankings can change as votes and endpoints change.[5][6]

## References

1. fal. "Introducing H3 Max by fal." August 2026. https://fal.ai/learn/devs/introducing-h3-max-by-fal
2. fal. "MiniMax H3 Max." Accessed August 28, 2026. https://fal.ai/minimax-h3-max
3. fal. "MiniMax H3 Max Text to Video API Docs." Accessed August 28, 2026. https://fal.ai/models/minimax/h3-max/text-to-video/api
4. fal. "H3 Max Image to Video API Docs." Accessed August 28, 2026. https://fal.ai/models/minimax/h3-max/image-to-video/api
5. Artificial Analysis. "Image to Video Leaderboard." Accessed August 28, 2026, at 05:37:52 UTC. https://artificialanalysis.ai/video/leaderboard/image-to-video
6. Artificial Analysis. "Text to Video Leaderboard." Accessed August 28, 2026, at 05:37:52 UTC. https://artificialanalysis.ai/video/leaderboard/text-to-video
7. MiniMaxAI. "MiniMax-H3" model card. Hugging Face. Accessed August 28, 2026. https://huggingface.co/MiniMaxAI/MiniMax-H3