MiniMax H3 C++/GGML Implementation Benchmarks: 10s Video in 119-130s on RTX 5090

Acceptable-Cycle4645 · reddit · 2026-08-08

The author shares progress on a C++/GGML implementation of MiniMax H3 and seeks performance baselines. Current benchmarks on RTX 5090 show 480x832, 10 steps, 10s video generation takes 119-130s with 24GB peak VRAM. Different memory-saving modes trade off speed: level 2 uses 11.4GB but takes 183s, level 3 uses 12.6GB and takes 163s. The author notes audio quality at 10 steps is better than the Python reference, and TTS word error rate is lower.

Original post →

More from coding & agent

coding & agent channel →