Kandinsky Video 5.0

A line‑up of lightweight (2B parameters), high‑speed models for text‑to‑video and image‑to‑video generation of up to 10‑second clips at up to 768×512 resolution — ranked top‑5 among open‑source models on the Text‑to‑Video Arena

• SFT models — highest generation quality after fine‑tuning on curated video and image data
• No CFG distilled models — 2× faster inference by removing classifier‑free guidance
• Distilled 16‑step models (Flash) — 6× speedup via Trajectory Segmented Consistency Distillation
• Pretrain checkpoints — for further fine‑tuning and research

A line‑up of high‑capacity (19B parameters) models for text‑to‑video and image‑to‑video generation of up to 10‑second clips at high resolution. Delivers state‑of‑the‑art visual fidelity, cinematic motion dynamics, and precise prompt adherence — ranked #1 among open‑source models on the Text‑to‑Video Arena

• SFT models — highest generation quality after fine‑tuning on curated video and image data
• Distilled 16‑step models (Flash) — 6× speedup via Trajectory Segmented Consistency Distillation
• Pretrain checkpoints — for further fine‑tuning and research