MiniMax H3 Max is a video-generation model developed by fal Research from the open-weight MiniMax H3 foundation. Announced on August 27, 2026, it combines additional post-training with an inference system optimized specifically for the new model. The headline result is unusually simple: fal reports that H3 Max can generate a five-second 768p video with synchronized audio in under three seconds.
That makes it faster than real time in the literal playback sense—the generated clip lasts longer than the time required to render it. For creators, the practical benefit is a much shorter prompt-test-revise loop. For developers, it creates opportunities for interactive video products that would be difficult to build around a model that takes minutes per result.
What Is MiniMax H3 Max?
MiniMax H3 Max is not a brand-new foundation model trained from scratch, and it is not an official replacement for every MiniMax H3 workflow. fal started with MiniMax H3’s released weights, added substantial new data during post-training, and focused on stronger prompt adherence, audiovisual quality, and aesthetics. Its research and inference teams also optimized the model and serving engine together rather than treating deployment as a separate final step.
According to fal’s launch report, the team continuously compared checkpoints through human preference studies. Speed optimizations were retained only when they preserved the desired quality gains. This matters because a “turbo” model can otherwise become faster simply by reducing sampling work and accepting weaker output.
MiniMax H3 Max keeps important H3 capabilities, including native synchronized video and audio. Its public endpoints currently focus on text-to-video and image-to-video generation, with optional first-and-last-frame control available through the image workflow. Output is offered at 480p or 768p, and 768p is the default resolution around which the service is tuned.
MiniMax H3 Max vs MiniMax H3
The best choice depends on whether a project values speed or breadth. The original MiniMax H3 is a 33-billion-parameter, general-purpose multimodal video model. It can interpret text, images, video, and audio in one context; generate native stereo audio; and produce clips up to 15 seconds. Its complete workflow also supports 2K regeneration, multiple reference assets, motion transfer, and video editing.
Capability | MiniMax H3 Max | MiniMax H3 |
|---|---|---|
Developer | fal Research, based on H3 | MiniMax |
Main goal | Speed, prompt adherence, aesthetics | Broad multimodal generation and control |
Public output resolution | 480p or 768p | 768p base output, with up to 2K regeneration |
Main public workflows | Text-to-video and image-to-video | Text, first/last frame, multimodal reference and editing workflows |
Audio | Native synchronized audio | Native 32 kHz stereo audio |
Reported five-second latency | Under three seconds at 768p | More than one minute in fal’s launch comparison |
Best fit | Rapid iteration and interactive experiences | High-resolution or reference-heavy production |
H3 Max is therefore not automatically “better” in every situation. Choose Max when short feedback cycles and strong instruction following matter most. Choose the original H3 when the workflow needs 2K output, several image/video/audio references, or advanced editing. MiniMax’s official H3 model card documents support for up to nine images, three video clips and three audio clips within its reference workflow.
MiniMax H3 Max Speed and Benchmark Results
fal reports approximately 2.4–3 seconds of wall time for a five-second 768p H3 Max generation. That represents roughly 35 times the throughput of the official H3 endpoint and, in fal’s tests, about 15 times the speed of models with comparable quality.
fal-published evaluation | H3 Max result |
|---|---|
Five-second 768p generation | Approximately 2.4–3 seconds |
Throughput versus official H3 endpoint | Approximately 35× |
Human preference evaluation | Ranked first for overall preference, prompt understanding and aesthetics |
Evaluation scope | Head-to-head comparisons with 12 video models using Bayesian Elo and 95% confidence intervals |
These are vendor-published results, so they should be treated as launch evidence rather than a guarantee for every prompt. End-to-end latency can also include queueing, safety checks, prompt expansion, uploads and delivery. In other words, “faster than real time” describes the relationship between render time and video duration; it does not mean that every user interaction has zero delay.
fal illustrated the distinction with H3 Max Live, an experimental infinite broadcast directed by chat. Viewers could enter !prompt, and the requested scene appeared within seconds. This was real-time inference and orchestration—not live training of the model’s weights. The demonstration shows why rendering faster than playback matters: the system can prepare upcoming generated footage while the audience is still watching the current segment.
MiniMax H3 Max Pricing
As of August 31, 2026, fal lists promotional launch pricing of $0.025 per generated second at 480p and $0.04 per second at 768p. The official H3 Max endpoint states that standard pricing after the promotion is $0.05 per second at 480p and $0.08 per second at 768p.
Resolution | Launch rate | Standard rate | Standard cost for 5 seconds |
|---|---|---|---|
480p | $0.025/second | $0.05/second | $0.25 |
768p | $0.04/second | $0.08/second | $0.40 |
These are fal platform prices and can change. They should not be interpreted as future SeeAPI pricing.
MiniMax H3 Max API Availability on SeeAPI
H3 Max already has public text-to-video and image-to-video API endpoints on fal. It is not yet available through SeeAPI at the time of writing. SeeAPI plans to move quickly to integrate MiniMax H3 Max as soon as the model is available for the platform’s production integration, giving users a convenient way to access it alongside other leading media models. Availability, supported parameters and SeeAPI pricing will be announced when the integration is ready.
Until then, developers can explore the existing MiniMax H3 API on SeeAPI for multimodal reference inputs, native audio and high-resolution video workflows.
Is MiniMax H3 Max Worth Watching?
MiniMax H3 Max is notable because its biggest improvement changes the product experience, not just a benchmark number. A five-second clip arriving in less than three seconds makes AI video feel interactive, supports rapid creative iteration and opens the door to chat-directed broadcasts, responsive storytelling and other continuously generated experiences.
The trade-off is equally clear: today’s Max endpoints prioritize 480p/768p speed and narrower generation workflows, while MiniMax H3 remains the stronger option for 2K, complex multimodal references and editing. H3 Max is best understood as a focused, post-trained H3 variant built for speed without intentionally giving up prompt adherence or visual appeal—and a model SeeAPI intends to support quickly once integration access is ready.
