Artificial intelligence Photo: VCG

Several Chinese artificial intelligence (AI) companies on Friday announced their latest tech achievements, with DeepSeek rolling out a public beta application programming interface (API) for its flagship V4-Flash model, MiniMax launching its multimodal H3 model and ByteDance releasing its new Seedance 2.5 video creation model.

Chinese AI companies are iterating at a rapid speed, making continuous breakthroughs in frontier areas such as video generation, multimodal understanding, and agent capabilities. Chinese experts said that this progress is a natural outcome of years of sustained investment, noting that in the AI field, openness, inclusiveness and cooperation are vital to achieve win-win results.

DeepSeek said on Friday that it has massively upgraded its agent capabilities, with benchmark scores now far surpassing the V4-PFro-Preview. The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex. It added that the official release of DeepSeek-V4-Pro will be coming soon.

MiniMax released its H3 model on Friday. This is a general-purpose multimodal generation model that supports unified understanding of multimodal contexts composed of text, images, videos, and audio. It can natively generate dual-channel audio-video content, with a maximum of 15 seconds at 2K resolution.

"We are capable of delivering the best performance-to-price ratio in the industry. By default, we provide 2K resolution, with a per-second price of less than one-third that of mainstream models at 2K resolution, and less than half at 768P resolution," the company noted.

For a long time, closed-source models have dominated the video generation field, with iteration speed and ecosystem openness lagging behind large language models. MiniMax said it plans to promote the development of the open-source community, accelerate compatibility with domestic chips and ensure users can customize their own versions.

After Seedance 2.0 gained wide attention both at home and abroad, ByteDance on Friday released the new Seedance 2.5 video generation model. This version extends the maximum single video generation length from 15 seconds to 30 seconds and supports multi-round extension. It also further enhances multimodal reference capabilities, allowing users to input up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single generation.

Seedance 2.5 has also begun to move into more specific production stages. According to the company, the model can generate high-quality synthetic video data to help train robots' perception and manipulation capabilities. It can also be used for industrial simulation, process training, and equipment demonstrations. In autonomous driving scenarios, the model can simulate low-frequency situations such as extreme weather and complex road conditions, providing additional samples for system testing and training.

"The concentrated release of these achievements is no coincidence. After years of accumulation, domestic AI models have gained advantages in performance, open-sourcing, and cost, ushering in a period of concentrated breakthroughs," Hu Qimu, a professor at the Maritime Silk Road Institute of Huaqiao University, told the Global Times on Friday.

Hu noted that China has a vast application market and a well-developed industrial system that can support multi-scenario implementation. In addition, national initiatives such as "AI+" further lower the threshold for industrialization and accelerate the pace of deployment.

Some senior US officials have recently said that they will launch investigations into Chinese AI companies over so-called "distillation" of leading US frontier models, and may impose sanctions on Chinese enterprises on grounds such as "stealing US intellectual property."

China's Ministry of Commerce slammed the US moves, which lack factual basis and legal justification, noting that innovation is not the exclusive domain of any single country. "Chinese AI companies have long been committed to advancing fundamental research, balancing technological self-reliance with open cooperation," a spokesperson noted.

At its core, competition in AI is a competition of ecosystems, rather than a contest over individual technology. Artificially building "small yards with high walls" and converting technological advantages into technological hegemony may yield commercial or political gains in the short term, but in the medium to long term it will lead to self-isolation and ultimately result in being excluded from a healthy ecosystem, Hu said.

Hu noted that in the face of smears and pressure from foreign countries, Chinese tech companies firmly adhere to the path of independent innovation, continuously strengthen their open-source efforts, and contribute to the global development of AI.

According to CCTV News, open-source models developed by China now account for 41 percent of global downloads, surpassing the US to rank first in the world. Moreover, the top five models by call volume on OpenRouter in July were all Chinese models, ranked as follows: Xiaomi MiMo-V2.5, DeepSeek V4 Flash, Tencent HY3, Zhipu GLM-5.2, and DeepSeek V4Pro. 

"China has become a major supplier in the global open-source AI ecosystem, significantly expanding the reach of advanced AI technologies and accelerating innovation and popularization of AI worldwide. Only by embracing openness, inclusiveness, and cooperation can true win-win results be realized," Hu noted.