Fal Speeds Up Open Source Video Model 35x

Engineers at fal reworked MiniMax's open source H3 video model to cut its generation steps dramatically, and say Hollywood clients have become one of the platform's fastest-growing user groups as a result.

MiniMax released its open source video model, H3, last month, and the team at fal decided to see how much faster they could make it run. According to fal's Gorkem Yurtseven and Batuhan Taskaya, the rebuild reduced the number of steps the model needs to take to produce a video, resulting in a version that runs 35 times faster than the original release. The pair discussed the project in a recent conversation shared by a16z, walking through how they approached optimizing an open source model rather than building one from scratch. Cutting down the steps required to generate a clip is a common lever for speeding up diffusion-style video systems, since fewer steps generally means less compute time per generation, though it can risk quality tradeoffs if not handled carefully. What stood out in the discussion was less the technical mechanics and more the demand side: Yurtseven and Taskaya said that after the faster version was built, Hollywood became one of fal's fastest-growing customer segments. That suggests studios and production-side users are actively hunting for open source video infrastructure that can be optimized for speed rather than relying solely on closed, proprietary systems. The full exchange, including more detail on how the pair approached the optimization work and what studio clients have been asking for since, is available via a16z's post on X. Neither MiniMax nor fal detailed further specifics on quality benchmarks or which studios are involved, but the episode is another sign of open source video models becoming a real battleground for infrastructure providers competing to serve professional production customers.