fal Launches H3 Max: 5-Second Video in Under 3s, ~35x MiniMax H3 Throughput
MiniMax_AI · x · 2026-08-28
fal launched H3 Max, a post-trained version of the open-weights MiniMax H3 co-designed with a custom inference stack for stronger prompt adherence, visual quality, and speed. In fal's human preference evaluations, H3 Max ranks #1 in overall quality, prompt understanding, and aesthetics — while generating a 5-second video in under 3 seconds, roughly 35x the throughput of the official MiniMax H3 endpoint and on average 15x faster than anything with comparable quality, challenging the assumed tradeoff between quality and inference speed in generative video.
The team continuously tested for quality while pushing speed, keeping only optimizations that preserved post-training gains; post-training added substantial new data focused on prompt adherence and visual quality, keeping the base model's core capabilities intact at very low latency. The post notes this required two capabilities that rarely sit under one roof: frontier model research and deep inference optimization/kernel work.
MiniMax congratulated the launch, saying H3 Max combines SOTA video quality with a step-change in generation speed, making high-quality video generation practical across broader real-world applications, and that the two teams have worked closely since day one.
Related event: fal's post-trained MiniMax H3 Max tops video leaderboard with 50x speedup(23 posts)→
More from Infra
- DFlash 2 Introduces Block-Diffusion Speculative Decoding to Speed Up GLM-5.3-Flash — songhan_mit · 2026-08-28
- M3 Max achieves 70 tok/s locally with 20GB RAM free after 32k task — mayfer · 2026-08-28
- Dev runs 125B Qwen model locally on M3 Max at 70 tok/s via MLX — mayfer · 2026-08-28
- NVIDIA launches Mesh open compute network to aggregate idle GPUs for AI — nvidia · 2026-08-28
- AI automated research finds numerical bug in vLLM/SGLang backend — josh_tobin_ · 2026-08-28
- Hot Chips 2026: inference chips enter an "era of ferment" with divergent bets — BenBajarin · 2026-08-28