We’ve been tweeting Sol-H3 in pieces — 8×GB200, Spark, 5090, Sol-Attn, the two-stage recipe.
Today the write-up is in one place.
Sol-H3: a tech report on full-stack MiniMax-H3 inference, cloud → edge.
Cross-res two-stage (no VAE round-trip) + Sol-Attn + an RSI loop that searches kernels and layouts.8×GB200:
5s of 768p in 1.43s (~3.5× realtime)
1× Spark: ~56s, HBM-resident
1× 5090: ~40s E2E
Up to ~30× vs the matched full-res baseline, ~20% less memory.
📄
🌐
💻
顯示更多