fal dev spills: MiniMax H3 'quality' mode matches balanced in 99.9% of cases but adds 60-120s
isidentical · x · 2026-09-04
fal developer isidentical revealed internals of MiniMax H3 Max Turbo's 'prompt expansion mode': the quality setting invokes MiniMax's proprietary prompt expansion model, taking over 60-120 seconds, yet in their evals quality performs on par with the default balanced mode in 99.9% of cases. He suggested removing the option entirely since 'so many people fall for it,' and recommends balanced for actual speed.
This responds to altryne's live test where Turbo generated slower than Max with poor prompt adherence. The post includes fal's H3 Max Turbo API page and a sample expanded prompt output.
Related event: MiniMax H3 Max Turbo tested slower than Max; quality mode called useless(3 posts)→
More from Multimodal
- AI film The Miracle to premiere Saturday directly on X — Kyrannio · 2026-09-04
- burnz releases AI-produced original single "Don't Make a Sound" — EddyVGG · 2026-09-04
- Meta's first image model Muse Image debuts #4 in image editing at just $0.01 per image — ArtificialAnlys · 2026-09-04
- One Prompt, One 3D Render: Astra Turns a Sentence into a Product Visual — iansilber · 2026-09-04
- The Generative Burrito Test puts 24 image models to work recreating one real burrito photo — generativist · 2026-09-04
- Immunologist one-shots a 5-minute T-cell lecture video from a single prompt with GPT-6 — DeryaTR_ · 2026-09-04