MiniMax H3Max generates 5s 768p video in under 3s, tops video benchmarks
赛博禅心 · wechat · 2026-09-01
MiniMax H3Max, post-trained by Fal with new data, verifiable RL, and inference acceleration, achieves 35x the throughput of base H3: a 5-second 768p audio-video clip in under 3 seconds and 15-second videos within 15 seconds — faster than playback, crossing the realtime threshold. It ranks #1 on ArtificialAnalysis and DesignArena image-to-video leaderboards. In three weeks H3 passed 24M downloads and 300+ derivative models, making it 2026's most downloaded model.
Notable uses: a Fal engineer wired H3Max into a Twitch stream where chat prompts change the video live; Pieter Levels launched infiniteslop.ai, a 24/7 AI livestream stitching viewer prompts. FastVideo, NuvaLab, and NVIDIA FastGen released FastH3, compressing 49 transformer forwards to 4 with 90% sparsity VSA — 15s 768p video in 13s on 8x B200, up to 14x speedup on a single Blackwell GPU — with weights, LoRA, and inference stack open-sourced.
The author also deployed H3 (and H3Max for a limited time) on AGIBar, free and unlimited, with a working curl command, and shared plans for a dedicated machine room in Beijing with 175kW reserve power.
More from Infra
- MiniCPM-V 4.5 MoE reportedly hits 13+ tokens/s on iPhone with just 3GB memory — heyshrutimishra · 2026-10-16
- Solo-built Vulkan inference engine LaForge hits ComfyUI-level SD1.5 speed: 512px in 0.8s, 402MB container — Ichidan256 · 2026-10-16
- Trading 2 RTX 5090s for 4 R9700s: does doubled VRAM pay off? — poofph · 2026-10-16
- Running PolyStrata + Cline coding agent on a $0-fancy laptop at ~90 tok/s — rpwoerk · 2026-10-16
- TensorFold 1.0.12 runs 120B vision model on 1-2 DGX Sparks, adds Mac image input — HankYeomans · 2026-10-16
- vLLM brings frontier model local deployment to NVIDIA RTX Spark PCs — vllm_project · 2026-10-16