fal's H3 Max: RL post-training cuts 5s video generation from 120s to under 3s
isidentical · x · 2026-09-16
fal Research launched H3 Max, a post-trained video model claiming #1 in overall quality, prompt understanding, and aesthetics against leading video models on first- and third-party evals, while generating a 5-second 720p video in under 3 seconds (50% off for a week).
Head of engineering isidentical explained the approach at MTS:
- Built on the open-source MiniMax H3 video base model, with reinforcement learning applied to verifiable concepts like video editing and prompt adherence
- Generation time cut from 120s to under 3s — nearly 2x real time — while improving quality over the open-source version and many frontier closed models
A notable release for anyone tracking real-time video generation.
More from Multimodal
- Midjourney wallpaper recipe: sref + profile mix for eerie film-blur portraits — michaelrabone · 2026-11-02
- Cavallaero Teaser: A Painterly Rider Races Through a Golden Forest — CommunicationTop4244 · 2026-11-02
- Anima's snapref Offers a New Reference-Based Fix for Character Consistency — Intrepid_Swim_8669 · 2026-11-02
- Pencil-textured folding fan: surreal product concept render goes viral — Delahuntagram · 2026-11-02
- Free Midjourney SREF gallery adds 1,200+ handpicked style codes overnight — tisch_eins · 2026-11-02
- Adobe's EditAgent beats prior image-editing agents but needs ~10x the compute — emmanuelvivier · 2026-11-02