MiniMax H3 Max

MiniMax H3 Max is a video generation model aimed at turning short prompts into ready-to-publish motion clips. Provider pages describe its output as polished and prompt-faithful, with generation that runs faster than real time, making it well suited to social-length edits, product teasers, and concept motion that needs to be iterated quickly rather than produced at feature-film scale. Practical workflows can start either from a written description or from a still image, with platforms accepting a reference frame alongside a core prompt to steer composition and motion.

On inference surfaces, the model is exposed as a standard video-generation task with configurable width, height, aspect ratio, duration, and seed controls, and outputs are delivered as MP4, WEBM, or MOV files. The same surface offers an async delivery option, webhook callbacks, and tunable safety settings, which suits pipeline integration in creative tools and short-form production stacks. Without an official model card in the supplied evidence, architecture, training data, and benchmark scores are not characterized here; the model's appeal rests on its prompt fidelity, reference-frame conditioning, and the speed advantage that lets creators move from idea to rendered clip in well under the clip's own running time.

Vercel AI Gatewayminimax/minimax-h3-maxminimax

Open provider page Provider docs

Quick Info

Provider: Vercel AI Gateway
Model key: minimax/minimax-h3-max
Release date: Aug 27, 2026
Last updated: Aug 27, 2026
Input modalities
Output modalities
Capabilities

Limits

Output tokens: 0 tokens
Context window: 0 tokens

Latest news about MiniMax H3 Max

Vercel AI Gateway

Coverage

MiniMax H3 Max is a post-trained variant of MiniMax's open-weight H3 video model, developed by fal Research and released on August 26, 2026, and listed on MiniMax's own API as MiniMax-H3-Max. Per the cited sources, it generates a 5-second clip at 768p in under 3 seconds of wall time (roughly 2.5 seconds of backend infe The CellCog breakdown adds that H3 Max supports both text-to-video and image-to-video generation with native audio in a single API call, clips up to 15 seconds at 768p or 480p, and image-to-video requests can use first-and-last-frame keyframing for precise control. It frames crossing the realtime threshold as the key c

Vercel AI Gateway

CoverageAnalysis

A technical deep-dive published on note.com describes fal.ai's release of a video generation model called "MiniMax H3 Max" on August 27, 2026. According to the article, H3 Max is built on MiniMax's open-weight H3 base model, with fal.ai performing post-training for cost and quality and co-designing the inference engine The article is a third-party, auto-translated secondary source rather than an official MiniMax or fal.ai announcement, so its benchmark and quality claims should be read as fal.ai-framed rather than independently verified. Attribution should preserve MiniMax as the base-model creator and fal.ai as the post-trainer and

Vercel AI Gateway

CoverageRelease Notes

fal announced H3 Max on September 1, 2026 as a post-trained video generation model built on the open-weights MiniMax H3, with fal Research performing the post-training and fal's inference team co-designing the inference engine. According to the supplied excerpt, H3 Max generates a five-second video in approximately thr The same announcement reports leading performance in independent evaluations: Design Arena ranks H3 Max first on its Image-to-Video leaderboard with an Elo of 1,341 (ahead of the official MiniMax H3 at 1,333, plus Seedance 2.5, FLUX.3, and Gemini Omni Flash), and Artificial Analysis ranks fal's H3 model first on its Im

Vercel AI Gateway

Coverage

Latent Space's AINews roundup covers the launch of fal's 'H3 Max Live', built by post-training MiniMax's H3 release and optimizing it for fal's in-house inference engine to reach about 35x the throughput of the official endpoint. The piece frames this as crossing an 'infinite video singularity' where generation runs fa The roundup also documents community demos built on top of H3 Max, including an infinite Twitch-style livestream where every frame is generated on demand from chat prompts, and an 'infinite AI-generated slop' interactive stream from levelsio. These examples illustrate the practical implication of sub-realtime generatio