fal

fal is a generative-media inference provider whose model pages and API are a standard commercial surface for other labs’ video models. In August 2026 it became a vendor in its own right: fal Research shipped H3 Max, a post-trained fine-tune of MiniMax’s open-weight H3, which debuted above the base model on Artificial Analysis’ image-to-video board while generating a 5-second clip in under 3 seconds. That combination — post-training a lab’s released weights and co-designing the serving stack around them — is the clearest instance so far of the infrastructure layer moving from distributing models to authoring them. See the AI Video Stack for full model coverage.