Real-time, steerable video generation from a 30B model on H100-class hardware makes interactive uses (agents or users steering a live stream) feasible. Access is closed beta only for now.
Key takeaways · AI-distilled
Reka built a real-time version of its 30B-parameter video model with NVIDIA: in closed beta it streams continuous video at 720p and 24 fps indefinitely, with each segment conditioned on the last frame and steerable by natural-language instructions.
The original model needed over 100 forward passes per segment once Classifier-Free Guidance was counted, which Reka says put real-time generation at 30B parameters far out of reach.
Reka judged training-free caching such as TeaCache and Cache-DiT modest and input-dependent, and instead used timestep distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition → via NVIDIA FastGen, folding guidance into the weights; it reports an 11.8x speedup and real-time output even on H100s.
Terms in this piece · Glossary
transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
