MiniMax H3 Max: what it costs to run a channel

Per second, then per show

fal bills per second of generated video with no subscription and no minimum. The table converts that rate into the units a channel operator actually plans around.

Launch pricing is 75% off and ends Sep 7, 2026 according to the source. Budget against the regular price; the discount has already changed once.

The ratio matters more than the latency

• fal states a five-second clip renders in about three seconds, roughly 35x the official MiniMax H3 endpoint throughput. An independent Design Arena benchmark put the speed-up above 50x at equal quality.
• A clip that plays for five seconds and renders in three leaves two seconds for moderation, encoding, and handing the file to the player. Drop below one rendered second per played second and the stream stalls.
• The public endpoints are asynchronous. The clock starts when a GPU picks up the job, not when you submit, so operators keep one or two clips of buffer and watch queue position.

Two resolutions, five to fifteen seconds

• Resolution is 480p or 768p per request. There is no 1080p or 2K option on fal even though upstream H3 supports 2K.
• A single generation runs five to fifteen seconds, five by default.
• Inputs: text-to-video, image-to-video with an optional end frame, and reference-to-video with several reference images.
• Controls include six aspect ratios, a prompt expansion mode (balanced about one second, quality up to thirty), a seed, and a safety checker that is on by default.
• Upstream H3 generates native stereo audio. fal’s pages do not document an audio option or charge, and whether hosted endpoints return sound is unconfirmed here.

Three things share the name

• H3 Max - The public post-trained model behind the three endpoints on this page.
• H3 Max Director - An autoregressive continuous variant with up to two minutes of context that powers fal.live and is not exposed to developers.
• H3 Max Turbo - A second public endpoint billed at exactly half price. fal publishes no speed figure; Live Classroom reports “2x faster and half the cost”.

A queue, not a stream

• Endpoint ids: minimax/h3-max/text-to-video, minimax/h3-max/image-to-video, minimax/h3-max/reference-to-video, plus minimax/h3-max-turbo/text-to-video.
• All use fal’s standard queue API: subscribe and wait, or submit, poll, and fetch. Webhooks work too.
• The official JavaScript client is @fal-ai/client with a FAL_KEY environment variable. fal’s landing page says the first five generations a day are free.
• There is no documented streaming endpoint. A public-API channel is a loop that keeps submitting the next clip while the current one plays.

H3 Max is not open source

• What is open is MiniMax H3, released Aug 5, 2026 as a 33-billion-parameter dense single-stream transformer with two CFG-distilled BF16 checkpoints.
• H3 Max starts from those weights and adds fal post-training focused on prompt adherence and visual quality. Training and inference run on NVIDIA GB200 NVL72 systems.
• The weights ship under the MiniMax H3 Community License, not Apache or MIT. Users in the United States, European Union, United Kingdom, and South Korea must file an application for rights.
• The release is partial: Context-IR, the full 2K workflow, and sparse-attention inference are held back. MiniMax’s official serving example uses four GPUs; Slop TV runs the model on two RTX 5090s at 352p.

How this category got here

• Jul 31, 2026 - MiniMax announces H3 with 15-second clips, 2K output, and native stereo audio.
• Aug 5, 2026 - MiniMax publishes weights and inference code under the H3 Community License.
• Aug 27, 2026 - fal releases H3 Max and posts the first stream generating faster than playback.
• Aug 29, 2026 - fal launches fal.live on H3 Max Director. Infinite Slop and Renoise Live go up the same week.