.@fal's Gorkem Yurtseven and Batuhan Taskaya on making an open source video model 35x faster, and what Hollywood wanted after they built it: Last month, MiniMax released H3, an open source video model. fal rebuilt it - they cut down the steps the model takes to make a video, rewrote the code under each stage, and got the GPUs to 70-80% of their theoretical ceiling instead of their usual 30-40%. No quality loss. Video now generates faster than you can film it. The models have gotten so cheap and fast that end users aren't even asking for improvements in either category anymore. The gap has moved to quality, or how closely the model follows the prompt. Hollywood wasn't a customer a year ago and is now fal's fastest growing segment. Studios love it for the little things - extending a shot, moving the camera, changing the lighting. Those edits land 80-90% of the time. fal is chasing 99.9%. In this conversation with a16z's Jennifer Li: 00:00 Intro 01:45 The first open model worth rebuilding 03:45 The industry ran out of compute in April 06:25 35x faster without losing quality 11:05 5s of video generated in 1.5s 14:25 The weekend fal dropped everything 16:15 3 viral projects nobody planned 18:20 Teaching a video model to remember 20:25 An hour of video that's consistent 23:30 2x the usage of other models in 3 weeks 26:35 The case for giving video away for free 29:00 Blender sketch in, finished shot out 30:40 Chasing 99.9% reliability 34:15 Hollywood, fal's fastest-growing customer 36:50 Why studios wouldn't touch it until now YouTube: @gorkem @isidentical @fal @JenniferHli

I tested 5,000, 8,000, and 10,000 step runs with the dataset. The captions mattered, adding an explicit matte toon style and descriptions of individual areas within each set helped anchor the look we wanted.

fal head of engineering @isidentical reveals how reinforcement learning turned H3 Max into a frontier video model that generates 5 seconds of video in under 3 seconds, down from nearly 2 minutes: "It's probably the only frontier video model from a quality and output and adherence perspective that can generate videos in real time." "We took this open source base MiniMax H3 video model, what we wanted to do was not just make it faster, but improve its quality ahead of anything else by implementing reinforcement learning for verifiable concepts like video editing and prompt adherence." "We were able to get the model from 120 seconds for a 5-second video generation, almost 2 minutes, to under 3 seconds for a 5-second video generation. Almost double real time, while improving its quality ahead of the open source version and ahead of many frontier closed source models." @fal

We celebrate one milestone by announcing another. Pickford x fal. Tonight we’re at the 78th Emmy awards where our project Whispers is a finalist for Outstanding Innovation in Emerging Media Programming. It also marks the beginning of an exciting collaboration with the team at @fal. fal is working closely with us to post-train custom models on our own in-engine art for Whispers, our interactive murder mystery where the audience shapes the story live. We could never visualize a story this way before and we’re excited to show you what comes next.