MiniMax introduces the H3, an omni‑modal video model that generates 15‑second 2K clips with native stereo audio, priced at $0.13 per second and ranking #1 in video editing on Artificial Analysis. It uses a newly built H3‑VAE tokenizer that provides a 4× gain in effective sequence length, enabling affordable 2K output without upscaling 1080p.
Read original
reddit/r/machinelearningnews