Pika speeds up video generation 60% with pure inference engineering in H3 Flex
ghumare64 · x · 2026-11-02
Pika released H3 Flex, now the #1 open-weight model on DesignArena's text-to-video board, with exportable audio stems, up to 4K resolution, and a 15-second scored clip in as little as 7 seconds.
The headline: a 60% speedup came entirely from inference engineering, without touching model weights or reducing the six denoising steps. Built on MiniMax H3 (already distilled from 50 steps to 6 via distribution matching distillation), the team attacked GPU bottlenecks: caching timestep-dependent AdaLN computations, kernel fusion, FlashAttention-4, block-sparse attention, and compressed inter-GPU communication.
The pipeline adds three post-training stages—SFT, six-step PDMD distillation, and post-distillation RL (ReFL plus offline flow-DPO adapters merged for inference). Human evals led on anatomy, complex composition, and audio sync.
More from Infra
- Researcher hoards nearly 1TB of DRAM, refuses to sell a stick to data centers — kevinnbass · 2026-11-02
- Self-Hosting LLMs at a Law Firm: Infra and Data Governance Harder Than the Model — EducationalGroup7245 · 2026-11-02
- Predictions: Opus 5.5-tier models at V4.1F inference cost within a year — teortaxesTex · 2026-11-02
- Google energy exec: space data centers cost up to 100x more than Earth-based ones — draecomino · 2026-11-02
- First convergence theory for PipeDream pipeline parallelism, with explainer video — peter_richtarik · 2026-11-02
- Modal CEO on who wins inference, open-source peak, and RL's real limits — 20VC · 2026-11-02