Sol-H3: MiniMax-H3 video generation faster than playback at 1.653s per 5s clip on 8x B300
MiniMax_AI · x · 2026-09-08
NVIDIA's Sol team and MiniMax released Sol-H3, their fastest end-to-end MiniMax-H3 inference stack. On a single 8x B300 Blackwell system it generates 5 seconds of 1344x768 video with stereo audio in 1.653 seconds — faster than playback. Versus 50-step Base H3 Dense on the same system: 5s clips go 18.250s→1.653s (11.04x), 10s 50.660s→3.732s (13.57x), 15s 99.513s→6.612s (15.05x). It scales across GPU counts (4x B300: 2.918/6.993/12.542s for 5/10/15s), reaching up to 15.54x speedup. Fully open-sourced and available via the Reactor API sandbox.
More from Infra
- Running MiniMax H3 video gen on 4GB VRAM: blurry hands at 512p or 6+ hours per clip at native res — unbenannt1 · 2026-09-08
- Celesto AI launches CelestoFS, petabyte-scale durable workspaces for AI agent sandboxes — aniketmaurya · 2026-09-08
- Nvidia's stakes in its own chip buyers hit $99B, up from $7B a year ago — GaryMarcus · 2026-09-08
- Pairing a decade-old RX 480 with RX 7900 GRE boosts llama.cpp MoE inference 36% over single GPU — tabletuser_blogspot · 2026-09-08
- Extrapolating OpenAI's plots: agents may eat 25% of compute within 9-12 months — joshua_clymer · 2026-09-08
- User runs Pinokio with Wan2GP, MiniMax H3, Viggle-Animate on DGX Spark — cocktailpeanut · 2026-09-08