Reka and NVIDIA unveil real-time 30B video model: 720p at 24fps, 11.8x faster on one H100
RekaAILabs · x · 2026-09-03
Reka AI, together with NVIDIA, has built a real-time version of its video generation model: 30B parameters, steerable mid-stream via natural language, outputting 720p at 24fps.
It runs 11.8x faster on a single H100 with no major quality drop in blind evals. The company frames it as a step toward "World Language Action Models" — AI that acts within the physical world, not just generates it. A closed beta is open on request.
More from Infra
- Nina Schick: the AI buildout's real bottleneck is electricians and plumbers, not physics — NinaDSchick · 2026-09-03
- Dan Jeffries Dreams of a 5TB DGX Station Under His Desk Running Sovereign AI — Dan_Jeffries1 · 2026-09-03
- Weaviate 1.38 Ships Boost API for Query-Time Rescoring When Hard Filters Are Too Strict — philipvollet · 2026-09-03
- HUMAIN and AWS expand AI partnership with up to 50MW of capacity by 2028 — HUMAIN · 2026-09-03
- "The Biggest Danger Isn't AI, It's Getting Left Behind": Compute and Data Sovereignty Argument — eliano · 2026-09-03
- OpenAI, Cursor, Cloudflare all go down at once, sparking upstream outage speculation — op7418 · 2026-09-03