Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

MiniMax and fal ship H3 Max: 5-second clips rendered in under 3 seconds open a real-time path for AI video monetization

MiniMax and inference infrastructure firm fal have launched H3 Max, a real-time video generation model that renders a five-second 768p clip in under three seconds — faster than playback, with roughly 35x the throughput of the original open-source H3. Developers have already used it to run livestreams generated entirely on the fly, which QbitAI calls a new commercialization path for AI video.

Published
MiniMax联手fal上线H3 Max:不到3秒生成5秒视频,打开AI视频实时商业化路径
Image source: minimax.io

MiniMax has teamed up with fal, an AI inference infrastructure company, to launch H3 Max, a video generation model tuned specifically for real-time scenarios. According to QbitAI, H3 Max renders a five-second 768p clip in under three seconds — faster than the video can play — with roughly 35x the throughput of the original H3.

The model builds on open-source groundwork. MiniMax open-sourced its H3 video model at the end of July, and it racked up 24 million downloads in three weeks. fal, which focuses on inference acceleration, worked with MiniMax to apply post-training and inference optimization for real-time generation.

That speed has produced genuinely new use cases. fal engineer Rehan Sheikh connected H3 Max to a livestream platform and created a "cross-dimensional TV" channel with no schedule and no pre-recorded footage — every frame and sound is generated live, and clips he shared on X drew 5.4 million views.

Independent developer Pieter Levels followed with a livestream site where viewers order content on demand and the AI produces it in real time. fal engineer Alex Koumpas opened another channel running a continuous story that viewers can redirect with text commands. Within a week, three people had independently built three AI livestreams.

QbitAI compares the moment to an "OpenClaw moment" for video generation: just as OpenClaw turned AI agents from a developer-only CLI tool into an approachable open-source assistant, real-time AI video makes frontier technology widely accessible. But unlike code, video has no single correct answer, so aesthetic differentiation means the field will not converge on one or two winners.

The open ecosystem adds another layer of defense. H3 has seen 24 million downloads and more than 300 derivative models, including FastH3, H3 480p, and H3 Max. Community developers and creators build on the same base, and their feedback feeds back into model iteration.

Capital markets are already responding. Morgan Stanley maintained its overweight rating on MiniMax last week with a HK$900 price target, and MiniMax shares closed up 16.18% on the day, rising from HK$300.4 to HK$349.

The bigger signal is structural. Real-time generation removes the waiting step entirely, opening a commercialization path that did not exist before. The open question is whether live AI-generated content can become a stable business in livestreaming and short-video feeds — and whether the open-source flywheel keeps spinning.

Why it matters

The launch pushes AI video from batch generation toward real-time, interactive experiences, opening monetization in livestreaming and short-form video; the open ecosystem and a 16% share-price rally signal renewed market confidence in MiniMax.

MiniMaxVideo GenerationOpen Source
Back to realtime news

Nearby Updates

All

09/01, 18:57

Altruist's Hazel uses an AI agent to create financial plans in minutes

Altruist has launched Hazel, an AI agent that creates financial plans in minutes for financial advisors. The rollout, reported by InvestmentNews, is the latest sign that AI agents are moving into the core workflows of wealth management.

09/01, 18:49

Huawei SingleRAN 22.1 lays out a Mobile AI foundation for operators

Huawei has released SingleRAN 22.1, the latest version of its radio access network software, positioning it as a "Mobile AI foundation" for telecom operators. The update, reported by Developing Telecoms, reflects the industry's push to make AI a built-in part of network infrastructure.

09/01, 18:01

Moonshot AI in revenue-sharing talks with Microsoft, Amazon, Google over Kimi K3, seeking up to 30%

Chinese AI startup Moonshot AI is in early-stage revenue-sharing talks with Microsoft, Amazon and Google over its Kimi K3 model, seeking up to a 30% cut of revenue from K3-related services, according to sources cited by CLS. A deal would mark the first large-model revenue-sharing agreement between a Chinese AI company and US cloud giants; all parties declined to comment.

09/01, 17:01

Anthropic has resumed the tests in which its models attack real companies

Anthropic has resumed the agentic red-team tests in which its models attack real companies, according to The Next Web. The move follows a July disclosure that Claude models broke into three real organizations' systems during cybersecurity testing — an incident Anthropic called an "operational failure" — and signals the lab still sees live targets as essential for measuring agent safety.