Veda Sparse ships 10% keep-ratio block-sparse attention accelerator for MiniMax H3 T2VA

linoy_tsaban · x · 2026-09-27

The Veda Sparse team released a preview checkpoint on Hugging Face that accelerates MiniMax H3 text-to-video generation:

• A 275M-param fp8 tile-score predictor selects, per layer and per head, which 128-token key tiles each query tile attends to, so attention runs block-sparse at a 10% keep ratio instead of dense
• Targets 8 denoising steps (8 NFE); the preview checkpoint was trained 600 updates on 5.17s clips only, but generalizes to 10.1s and 14.4s geometries
• fp8 storage halves download/load while top-k selection keeps accuracy unaffected; weights and tile plans ship in one file to prevent silent mispairing
• Includes deployment guides for Ada (SM89), Hopper (SM90) and Blackwell (SM120)

Original post →

More from Multimodal

• Gemini Live Avatars go GA: real-time talking avatars in 97 languages, with pricing — Sam Witteveen · 2026-09-27
• Opus 5.5 animates Pink Floyd's "Time" from a style ref and a few prompts — Hesamation · 2026-09-27
• 9 open-source video generation model families to know in 2026, from Wan2.2 to Mochi 1 — TheTuringPost · 2026-09-27
• Astra 6 vs Opus 5.5 for motion graphics: one takes 6 tries, the other 46 minutes — aziz4ai · 2026-09-27
• Shared prompt template: 'Wild Magic Transformation' image generation with vines, dragon scales and elemental flames — LudovicCreator · 2026-09-27
• Music LoRAs pick up steam: YuE2 + Two Steps From Hell LoRA writes 3-min cinematic anthems — linoy_tsaban · 2026-09-27

Multimodal channel →