MiniMax and fal.ai ship H3 Max, a faster-than-realtime video model behind a new interactive AI livestream.
The AMW Read
Extends known MiniMax H3 open-weight model with fal.ai's inference speed-up and a live interactive-broadcast demo, incrementally advancing open-weight video-gen distribution without new base-model capability from MiniMax itself.
MiniMax and fal.ai ship H3 Max, a faster-than-realtime video model behind a new interactive AI livestream.
fal.ai opened a free demo site for MiniMax's H3 Max video model, generating 480p, five-second clips with audio in about three seconds from text or image prompts; registered users get five free generations daily. H3 Max is fal.ai's speed-tuned version of MiniMax's H3, the open-weight multimodal video model MiniMax released July 31 with support for up to 2K, 15-second audio video and editing. fal.ai says H3 Max runs roughly 35 times faster than base H3 and generates 15-second clips in about 15 seconds — real-time speed. API pricing is $0.05 per second at 480p and $0.08 per second at 768p; fal.ai co-founder and CTO Gorkem Yurtseven said H3 Max's weights will also be released openly, without a timeline. To demonstrate the speed, fal.ai launched fal.live, a continuous AI-generated broadcast where viewer votes pick the next scene, powered by an experimental H3 Max Director variant that references up to two minutes of prior footage for continuity.
This is fal.ai's second H3 speed optimization in three days, following Friday's cut of five-second clip generation to under four seconds, and it lands as MiniMax expands its own compute footprint — the company raised its three-year Alibaba Cloud spending ceiling 220% to $1.2 billion after burning two-thirds of its 2026 budget by June, per AMW's prior coverage. Per the AI Market Watch index, MiniMax has raised $1.77 billion in total funding since its 2021 founding (coverage limit: ~5,000 companies tracked, not a census). The move keeps MiniMax's video strategy anchored in open-weight distribution plus third-party inference tuning rather than a single closed stack, with value shifting toward inference latency and cost per second.
For builders, sub-realtime generation with audio and start/end-frame control makes interactive, viewer-steered live formats a real product category, and the $0.05–$0.08 per-second pricing sets a concrete cost benchmark. For investors, the fact that an outside inference specialist — not MiniMax itself — is driving the speed gains suggests margin and distribution power in generative video may concentrate at the serving layer as much as with the model owner.
#MiniMax #falai #VideoGeneration #GenerativeMedia #OpenWeightAI #AIInfrastructure
