Recent benchmark results show that fal’s H3 Max and H3 Max Turbo produce finished video clips with synchronized audio in less time than the clips themselves last.
SAN FRANCISCO, CA, UNITED STATES, September 8, 2026 /EINPresswire.com/ — fal has disclosed measured generation speeds for H3 Max and H3 Max Turbo, its post-trained versions of the open-weights MiniMax H3 video model. Across nine setups covering every accepted resolution and clip length, both endpoints produce video quicker than the video plays back. fal’s H3 Max is the fastest AI video generator of 2026!
These numbers represent generation time—the inference value each endpoint reports in its own response, reflecting the model’s GPU workload. The total wait for a caller adds queue time, prompt expansion, and encoding. Measurements were collected on September 8, 2026 for text-to-video with prompt expansion left at its default setting.
At 768p, the resolution the models are optimized for:
– 5 seconds of video generates in 1.54 seconds on H3 Max Turbo and 2.46 seconds on H3 Max
– 10 seconds generates in 4.29 seconds on Turbo and 7.55 seconds on H3 Max
– 15 seconds generates in 8.44 seconds on Turbo and 15.17 seconds on H3 Max
At 480p, a 5-second clip generates in 0.44 seconds on Turbo and 0.75 seconds on H3 Max, roughly 11 times and 7 times faster than the clip plays. At 1080p, a 5-second clip takes 2.33 seconds on Turbo and 3.12 seconds on H3 Max. Turbo returned every one of the nine configurations faster than real time, and H3 Max did so in seven of nine.
Video Length | Quality | H3 Max Turbo by fal | H3 Max by fal
5 second clip | 480p | .44 seconds | .75 seconds
10 second clip | 480p | 1.00 seconds | 1.72 seconds
15 second clip | 480p | 1.89 seconds | 3.14 seconds
5 second clip | 768p | 1.54 seconds | 2.46 seconds
10 second clip | 768p | 4.29 seconds | 7.55 seconds
15 second clip | 768p | 8.44 seconds | 15.17 seconds
5 second clip | 1080p | 2.33 seconds | 3.12 seconds
10 second clip | 1080p | 6.81 seconds | 8.89 seconds
15 second clip | 1080p | 13.56 seconds | 17.63 seconds
Picture and audio are produced in a single pass. A standard request yields a 5-second 768p clip at 1344 by 768 and 24 frames per second, with stereo audio embedded in the same file, so dialogue, room tone, effects, and music are described in the prompt rather than added in a secondary step.
fal attributes this speed to designing the model and inference engine as an integrated system, rather than training first and serving later. On the model side, fal Research introduced new training data and employed an in-house reinforcement learning framework to enhance prompt adherence and visual quality, retaining optimizations such as lower precision and reduced sampling steps only where human preference scores remained stable.
Regarding quality, Artificial Analysis ranks H3 Max first among image-to-video models with audio output in its Image to Video Arena, holding an Elo of 1200 ahead of the next model at 1192 and the base MiniMax H3 model at 1187. Those standings were recorded on September 8, 2026.
H3 Max is now available via fal’s serverless API across six endpoints: text-to-video, image-to-video, and reference-to-video on H3 Max; text-to-video and image-to-video on H3 Max Turbo; and H3 Max Director, a realtime model that maintains a streaming session rather than answering a single request. Clips range from 5 to 15 seconds at 480p, 768p, or 1080p, with aspect ratios from 21:9 to 9:16 and prompts up to 50,000 characters. Python and JavaScript SDKs are available, and the API can be called directly over HTTP.
Pricing is per second of output video with no minimums and no subscription. H3 Max costs 0.05 US dollars per second at 480p, 0.08 at 768p, and 0.16 at 1080p. H3 Max Turbo costs 0.025, 0.04, and 0.08 respectively. Both are offered at 75 percent off through September 14, 2026. Signed-in users receive five free generations per day of up to 15 seconds each, on a rolling 24-hour reset.
H3 Max and H3 Max Turbo are available at fal.ai.
About fal
fal is a generative media platform serving image, video, and audio models through a serverless API, running them on infrastructure it builds and optimizes itself. Learn more at fal.ai.
Bennett Heyn
fal
email us here
