fal speeds up MiniMax's open-source video model H3 by 35x, nearing real-time generation
In an a16z interview hosted by GP Jennifer Li, fal co-founder Gorkem Yurtseven and engineering lead Batuhan Taskaya detailed how the company optimized MiniMax's recently open-sourced H3 video model (H3 Max) to near real-time generation: 35x faster overall, pushing video generation into the real-time era.
Confirmed
- The optimization involved two main levers: reducing the number of sampling/generation steps, and rewriting low-level code across all stages for a full inference-stack refactor.
- GPU utilization was raised from a typical 30-40% to 70-80%, with one account noting near-theoretical peak utilization.
- Hollywood has become fal's fastest-growing customer segment.
Why it matters
- With MiniMax open-sourcing H3, fal's case shows that open-source video models can approach real-time generation purely through inference engineering—no new model required—suggesting inference-layer optimization can matter as much as the model itself.
- Doubling GPU utilization means the same compute can serve far more generation requests, directly affecting cost and scale, which is the backdrop for Hollywood's accelerating adoption of fal.
2026-09-18 ~ 2026-09-18 · 5 related posts
Primary sources
- [source] a16z podcast: how fal's H3 Max turns MiniMax's video model near real-time — a16z · 2026-09-18
- [source] fal made an open-source video model 35x faster — and Hollywood became its fastest-growing segment — a16z · 2026-09-18
- fal makes MiniMax's open-source video model H3 35x faster, GPU utilization hits 70-80% — astrange · 2026-09-18
2 near-duplicate retellings: isidentical · gorkem