NVIDIA Sol Engine Runs MiniMax-H3 on Desktop GPUs with Up to 4.52× Speedup
songhan_mit · x · 2026-08-06
NVIDIA's SANA team released the Sol Engine optimization for the omni-modal video model MiniMax-H3 (Hailuo 3.0), successfully bringing the massive 33B parameter model to desktop-class hardware.
By utilizing full-stack kernel fusion, cross-step caching, and training-free sparse attention (Sol-Attn), the solution uses the stock 33B checkpoint without any distillation, fine-tuning, LoRA, or offline calibration.
Benchmark Acceleration:
- 3.92× speedup on DGX Spark (480p, 24 FPS)
- 4.52× speedup on GeForce RTX 5090 (720p, 24 FPS)
This demonstrates the robust adaptability of agent-native optimization across various hardware and deployment settings, scaling from data centers down to edge devices.
Related event: NVIDIA Sol Engine Boosts MiniMax H3 Inference Speed by 3.95x(2 posts)→
More from Infra
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Nemotron 3.5 Bolt runs 16 concurrent streams on a single DGX Spark — NVIDIAAI · 2026-10-16
- Why your computer can't store 0.2: an IEEE 754 primer — Franc0Fernand0 · 2026-10-16
- Sabaki: MIT-licensed typed decision router runs on CPU in milliseconds — Fox3000foxy · 2026-10-16
- BitRun v0.2 runs 8B ternary LLMs on CPU in 2GB RAM with live wattage telemetry — The_VibePreneur · 2026-10-16
- MiniCPM-V 4.5 MoE reportedly hits 13+ tokens/s on iPhone with just 3GB memory — heyshrutimishra · 2026-10-16