- The open-source model targets the fast-growing AI video market with lower costs and broad chip ecosystem support
- Its release marks a new battlefront between closed commercial platforms and open AI models in video generation
MiniMax has released its latest multimodal model H3, entering a growing race in AI video generation with an open-source strategy aimed at challenging closed models from rivals including ByteDance’s Seedance.
The launch came on July 31, the same day ByteDance released Seedance 2.5, marking what industry observers described as a head-to-head showdown between two competing approaches to AI video development.
MiniMax H3 is a general-purpose multimodal generation model designed to process text, images, video and audio within a unified context. It can generate videos with up to 2K resolution, 15-second duration and native stereo sound, according to the company.
A key differentiator is MiniMax’s decision to release the model weights within days of launch.
On August 3, H3 was officially open-sourced, with 16 domestic and international chipmakers and developer ecosystems, including Huawei Ascend, Moore Threads, MetaX, Biren Technology, AMD and Intel, completing compatibility adaptations on the same day.
Two strategies emerge in AI video race
The release highlights a broader divide in the AI video market.
ByteDance’s Seedance currently holds more than 80% of China’s AI video model market, according to industry data.
Its advantage lies in product maturity, distribution and ecosystem integration. Seedance 2.5 extends single-generation video length to 30 seconds and focuses on turning AI video from an experimental tool into a production workflow, with stronger emphasis on maintaining narrative continuity.
MiniMax is taking a different route by combining open-source access with lower costs.
In Artificial Analysis’ audio-enabled text-to-video benchmark, H3 scored 1,242 points, ranking second globally and just four points behind Gemini Omni Flash, while outperforming Seedance 2.0. In video editing benchmarks, H3 ranked first globally with a score of 1,130.
Pricing is another competitive angle. MiniMax said H3 generation at 2K resolution costs about 0.8 yuan ($0.12) per second, less than one-third of many mainstream models.
Independent tests suggest the two models have different strengths. H3 performs better at maintaining visual consistency and overall cinematic coherence, while Seedance is stronger at following detailed directing instructions and shot-level requirements.
In one comparison involving a fantasy museum scene, H3 produced more consistent environments and character appearances, creating a more unified final video.
Seedance, meanwhile, was better at incorporating specific camera and storyboard instructions, though the final clips appeared less visually connected.
Can H3 revive MiniMax’s momentum?
H3 is seen as a critical release for MiniMax after its previous M3 model disappointed some investors and triggered a sharp stock selloff.
By open-sourcing H3, MiniMax is betting that the model can become infrastructure for developers and applications rather than just a standalone product.
The move is drawing comparisons with DeepSeek’s open-source strategy, which helped accelerate adoption of Chinese AI models globally.
However, turning technical performance into commercial success remains a challenge.
Although 16 ecosystem partners supported H3 at launch, a complete 2K generation workflow still relies on MiniMax’s API.
The company has not yet open-sourced its Context-IR module, meaning developers cannot fully reproduce the highest-quality generation process locally.
MiniMax also faces financial pressure. The company reported revenue of about $79 million in 2025 while spending $253 million on research and development, leaving profitability unresolved.
Open-sourcing may expand adoption, but monetization will likely depend on cloud services, enterprise support and customized solutions.
Competition in AI models is also becoming increasingly unforgiving. Large language models have seen performance benchmarks pushed higher by models such as Kimi K3, while low-cost alternatives such as DeepSeek V4-Flash have compressed pricing.
Companies that are neither the best-performing nor the cheapest risk being squeezed out of the market.
H3 has achieved a global lead in video editing benchmarks, but remains behind the top models in text-to-video and image-to-video rankings, where it places second and third respectively.
Citigroup responded with a “buy/high risk” rating and a target price of HK$533. Yet the longer-term significance of H3 may extend beyond its impact on MiniMax’s valuation.
For years, AI video generation has been dominated by closed models, with slower ecosystem development compared with large language models. By releasing H3’s weights, MiniMax is offering the industry an open experiment — one that developers can build on, validate or challenge.
The battle for H3’s commercial future has only just begun.
