fal releases H3 Max: #1-ranked post-trained video model at 30x inference speed
isidentical · x · 2026-08-27
fal Research introduced H3 Max, a post-trained video model built on a MiniMax base. It ranks #1 in overall quality, prompt understanding, and aesthetics on both first-party and third-party independent benchmarks (Artificial Analysis, Design Arena), while generating a 5-second 720p video in under 3 seconds — 30x+ faster inference than peers. It's 50% off for launch week, positioning it as the fastest, cheapest top-quality video model.
Related event: fal's H3 Max: Post-trained MiniMax H3 Tops Image-to-Video Leaderboards(29 posts)→
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15