(Yicai) Jan. 22 -- Chinese internet giants and artificial intelligence startups, including ByteDance, Moonshot AI, and DeepSeek, have revealed new large language model updates this week, claiming that some of them can solve math and coding challenges as well as OpenAI's latest O1 model.

On Jan. 20, DeepSeek introduced its DeepSeek-R1 model, and slightly more than an hour later, Moonshot AI launched its Kimi K1.5 model. On the same day, Shanghai-based MiniMax launched audio functions for its Hailuo AI text-to-video platform, while ByteDance's Doubao debuted its real-time voice model, which can act as an English teacher or storyteller for users.

Beijing-headquartered Moonshot AI claims that Kimi K1.5 has caught up with OpenAI's O1, which debuted last month, in mathematics, coding, and multimodal reasoning capabilities. Similarly, Hangzhou-based DeepSeek said that DeepSeek-R1 performs on par with O1 in tasks such as math, coding, and natural language inference.

Chinese programs are also less expensive. DeepSeek offers application programming interface services priced at CNY1 (US 10 cents) per million tokens of input and CNY16 (USD2.20) per million tokens of output, significantly lower than O1's rates of USD15 and USD60, respectively.

Strategy Divide

The latest updates by Moonshot AI and DeepSeek generally align with the inference direction of OpenAI's O1 series but may not use identical technical solutions, a research and development employee from a leading domestic LLM firm told Yicai.

Chinese companies are moving toward open-source models while OpenAI, the American developer of ChatGPT, maintains its proprietary source code strategy. For example, Moonshot AI has released its technical report regarding model training for the first time, while DeepSeek has revealed its model weights.

Jim Fan, senior researcher manager at Nvidia, praised DeepSeek's move toward shared AI development tools. "We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive - truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely," Fan posted on X on Jan. 20.

"DeepSeek-R1 not only open-sources a barrage of models but also reveals all the training secrets. They are perhaps the first OSS [open source software] project that shows major, sustained growth of an RL flywheel," Fan added, referring to a reinforcement learning system that gets rewarded from successful interactions, resulting in a positive feedback loop.

Moreover, social media and gaming giant Tencent Holdings yesterday released its Hunyuan3D 2.0, an open-source model that converts 2D pictures into 3D images, with expected applications in game development, e-commerce advertising, and industrial manufacturing.

Chinese tech behemoth Baidu has also recently begun testing a new AI application called Free Canvas, a whiteboard tool based on the firm's cloud platform and Baidu Wenku, a digital library that has evolved into an AI content generator. The number of monthly active users of Baidu Wenku has surpassed 90 million, said Wang Ying, vice president of the search engine operator.

Editor: Emmi Laine