MiniMax launched H3 on Friday, ranking first in video editing on Artificial Analysis benchmarks. The Shanghai firm is opening weights to undercut ByteDance's Seedance 2.0 and Google's gemini-omni-flash" class="entity-chip">Gemini Omni Flash.
Key facts
- H3 ranks #1 in video editing on Artificial Analysis
- Trails Gemini Omni Flash in text-to-video tasks
- Weights under MiniMax Community License, $20M revenue cap
- Released Friday, available via API and Hailuo AI
- Competes with ByteDance Seedance 2.5, launched same week
Open weights as a competitive lever
MiniMax launched H3, its newest multimodal video generation model, pledging to break closed-source "dominance" through open weights and competitive pricing — as rival ByteDance rolls out its latest Seedance 2.5 model. According to SCMP, H3 was currently the world's most powerful AI model in video editing, according to benchmark platform Artificial Analysis. However, it trailed Google's Gemini Omni Flash in text-to-video tasks, and ranked behind both ByteDance's Seedance 2.0 and Gemini Omni Flash in image-to-video capabilities.
Designed for commercial content creation across film, advertising, video games and product design, H3 is currently available via an application programming interface and MiniMax's consumer-facing Hailuo AI platform. Full model weights were slated for release "in the coming days", MiniMax said on Friday.
The licensing economics
While MiniMax's previous Hailuo video models were proprietary, it decided to open up H3's weights to "support the open-source community" and let users customise their own versions with "a broad range of AI hardware", MiniMax said, citing hardware compatibility as a "key consideration" in H3's design. Weights are the underlying parameters that encode an AI's intelligence.
"Closed-source models have long dominated video generation, with slower iteration and a less open ecosystem than fields like large language models," the company said in a blog post on Friday.
MiniMax would release H3's weights under the MiniMax Community License, which permits free non-commercial use and commercial use for organisations with under US$20 million in annual revenue with attribution requirements, according to Artificial Analysis. The $20 million revenue threshold mirrors the licensing structure that DeepSeek and other Chinese model vendors have used to capture developer mindshare without cannibalising enterprise revenue — a pattern that has worked in LLMs but is untested in video generation where compute costs for fine-tuning are higher.
What the benchmarks actually show
The Artificial Analysis rankings put H3 ahead of Seedance 2.0 in video editing, but the gap narrows in other modalities. ByteDance's Seedance 2.5, which generates 30-second video in one shot, was not included in the comparison — MiniMax is fighting a moving target. Google's Gemini Omni Flash remains the text-to-video leader, suggesting MiniMax's win is category-specific rather than a wholesale takeover of the video generation crown.
For commercial users, the API pricing and open weights create a genuine alternative to ByteDance's Jimeng platform. The question is whether H3's editing strength translates to production workflows, where Seedance 2.5's one-shot 30-second generation may be more valuable than superior edit quality.
What to watch
Watch for MiniMax's actual weight release in the coming days and whether the open-weights community produces fine-tunes that close the text-to-video gap with Gemini Omni Flash. Also track ByteDance's response on Seedance 2.5 pricing — if it cuts API rates, the open-weights advantage narrows.
Source: scmp.com
[Updated 02 Aug via scmp_tech]
MiniMax H3 supports native 15-second 2K dual-channel audio-video generation, a technical spec not previously disclosed, and prices video generation at 0.8 yuan per second (about $0.11), undercutting rivals on cost [per Pandaily]. This pricing detail sharpens the competitive picture against ByteDance's Seedance 2.5, which generates 30-second clips in one shot but at higher cost. The 2K native output also positions H3 for professional production workflows, potentially offsetting its text-to-video gap with Gemini Omni Flash.
[Updated 03 Aug via pandaily]
MiniMax has released H3's base weights on HuggingFace, revealing a 33B-parameter dense single-stream Omni-Transformer architecture with 16x spatial and 4x temporal compression VAE [per Pandaily]. The company also confirmed API pricing at one-third of Seedance 2.0's rate, a sharper cost advantage than previously reported. The model can run locally on two RTX 5090 cards, lowering the barrier for developers to fine-tune without cloud dependency. This release delivers on the promised open-weight rollout and gives the community a concrete target for closing the text-to-video gap with Gemini Omni Flash.
