Krea Realtime 14B: Real-Time, Long-Form AI Video Generation (www.krea.ai)

Krea Realtime 14B is a newly released 14-billion-parameter autoregressive video model that enables real-time, long-form text-to-video generation and interactive streaming. Distilled from Wan 2.1 14B using the Self-Forcing pipeline, it runs at about 11 fps with only 4 inference steps on a single NVIDIA B200 GPU and can return the first frames in ~1 second. Because it’s autoregressive (causal), users can edit prompts mid-generation, restyle streams on the fly, and direct video as it’s produced—capabilities that bidirectional diffusion models can’t provide. The core technical advance is scaling Self-Forcing to a 14B generator while solving exposure bias (the train/inference mismatch when the model conditions on its own imperfect outputs). The three-stage distillation combines timestep reduction, causal ODE pretraining with block-causal attention masks (bidirectional inside blocks, causal across blocks), and Distribution Matching Distillation (DMD) using pretrained bidirectional score models as critics—allowing a data-free, distribution-level loss. Engineering innovations include KV cache recomputation and KV cache attention-bias techniques to mitigate error accumulation, plus memory optimizations (dynamic KV cache management, gradient checkpointing, FSDP/ZeRO-3 sharding and dynamic freeing) that made training tractable at 14B scale. The result is a publicly available realtime video generator that materially advances interactive creative workflows and demonstrates practical ways to convert high-quality bidirectional video diffusion teachers into robust causal, real-time systems.