Summary
In this a16z podcast episode, the founders of Fal.ai sit down to discuss HDMax Turbo, their post-trained version of Minimax’s open-source H3 video generation model. They explain how systemwide optimizations let the model generate a 5-second video clip in as little as 1.5 seconds, at roughly half the cost of comparable systems, and why H3 represented the first truly next-generation open-source video model capable of using reference images.
The conversation digs into the engineering tradeoffs behind squeezing more “video pixels” out of each GPU, and how Fal’s distributed team pulled off a marathon multi-day development sprint to ship several projects at once. Listeners get a behind-the-scenes look at how three separate viral moments emerged organically — including an engineer live-streaming continuous H3Max generations on Twitch and creator Levelsio building his own infinite-streaming website — all without any coordinated marketing push.
The hosts and guests also explore the idea of “token market fit,” comparing generative video to coding agents as markets where a single power user can productively spend thousands of dollars in compute per month. It’s a useful listen for anyone tracking how open-source video models are being commercialized and optimized at the infrastructure layer.
📺 Source: a16z · Published September 17, 2026
🏷️ Format: Interview
