vLLM-Omni Renders MiniMax H3 Faster Than Playback: 10.1s Video in 8.7s
vllm_project · x · 2026-09-02
The vLLM project announced a collaboration with the FastVideo team (hao-ai-lab) and NVIDIA, treating MiniMax H3 serving as a systems problem—rendering a complete 10.1s MP4 with synchronized audio in just 8.7 seconds, i.e., faster than playback.
Technical highlights:
- Long-context attention + Fast Ulysses comms (NCCL SymmetricMemory)
- Fully fused DiT ops (RMSNorm+RoPE, SwiGLU, modulation)
- Parallel video/audio VAE decode across 8 GPUs
- 75% payload cut before D2H transfer, parallel H.264/AAC muxing
- FastVideo's 4-step FastH3 distillation removes the remaining bottleneck
FastH3 is an open-weight 4-step sparse-distilled MiniMax-H3 for synchronized video-and-audio generation, developed with Nuva Lab and NVIDIA.
Related event: vLLM and FastVideo Make MiniMax H3 Video Generation Faster Than Playback(7 posts)→
More from Infra
- Auction-based LLM routing paper finds fixed token pricing wastes 10%-71% in costs — _onionesque · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- SGLang Summit sponsors jump from 11 to 36 in one week — ying11231 · 2026-10-18
- Oracle CDS hits record 300bps, called first AI trade domino after Barclays sell downgrade — arthurcolle · 2026-10-18
- Agent Coding Workflow Would Cost ~$40k/Month on API Pricing, Dev Reveals — ChrisRackauckas · 2026-10-18
- PS5LM: llama.cpp port turns a jailbroken PS5 into a home LLM server (500+ GitHub stars) — wateriscoding · 2026-10-18