Open-source VDN speeds Minimax-H3 video generation 75-90x, faster than playback
anselm · x · 2026-09-03
Video Delta Net (VDN) introduces hybrid attention for near-lossless live text-to-video, accelerating Minimax-H3 by 75-90x — generating 14 seconds of 768p video in 11 seconds on 8x NVIDIA B200 GPUs. Checkpoints, training/inference code, and a technical blog are open-sourced.
More from Infra
- GLM-5.3-Flash beats DeepSeek-V4-Flash for writing and vision on 2× DGX Spark — kuhunaxeyive · 2026-09-03
- Jeff Dean's 2007 slide on hardware failures inside a datacenter resurfaces — SumitGup · 2026-09-03
- Google unveils 8th-gen TPU at Hot Chips: two chips per year, split inference and training designs — firstadopter · 2026-09-03
- Data center water consumption is a myth, says engineer: modern builds use closed-circuit cooling — GlenBradley · 2026-09-03
- Inference Engineering Is Just a Recipe: vLLM/SGLang, Replicas, Cache-Aware Routing — GabGarrett · 2026-09-03
- Databricks pitches agent-native data infrastructure, Lakebase Postgres at VLDB 2026 — matei_zaharia · 2026-09-03