The next DeepSeek to disrupt the global artificial intelligence industry could also come from China.

The country of 1.4 billion has seeded several promising AI startups and projects, while its leading internet players have spent years investing and developing the infrastructure to support such new ventures. With the DeepSeek experience casting doubt over the need for bleeding-edge expensive AI hardware from the likes of Nvidia Corp. — which the US has banned from sale in China — the prospects of China’s up-and-coming AI companies look to be improving.

In 2025, Alibaba Group Holding Ltd.’s Qwen platform has emerged as one of the leading contenders to match not just DeepSeek but also ChatGPT-creator OpenAI. On Thursday, it released its latest open-sourced QwQ: a model that it claims performs as well as DeepSeek with just a fraction of the data required. Another oft-cited player is ByteDance Ltd. and its Doubao bot. And Tencent Holdings Ltd.’s Yuanbao AI chatbot briefly topped the iPhone download rankings this week.

Here are the rising companies and services to keep an eye on, as Chinese businesses get creative in navigating trade curbs and developing more efficient AI models. The valuations are derived from Pitchbook data and previous Bloomberg News reporting.

Moonshot was founded by Tsinghua University assistant professor Yang Zhilin, who had earlier stints working on AI projects for Meta Platforms Inc. and Alphabet Inc.’s Google. Moonshot’s latest release leverages reinforcement learning to speed up and scale the training process. Announced on the same day as DeepSeek’s R1 reasoning model, it works similarly. Reasoning models use a chain-of-thought approach to finding and refining an answer to a query. Kimi k1.5 is designed for briefer responses and supports up to 2 million Chinese characters of context, which can comprise everything from prior conversations to user intent and circumstances.

Zhipu has released a series of apps and tools, including an autonomous agent that can simulate a user’s actions on a smartphone or a web browser. Its speech AI model includes simulations of human tone, emotions and imperfections, such as filler words, pauses and the occasional catching of breath. The company was blacklisted by the US Department of Commerce in the final days of the Biden administration, further narrowing its access to American AI chips. The elite Tsinghua University credited Zhipu’s success to its alumnus Tang Jie, whose team spearheaded the company’s model development. Tang also mentored Moonshot’s Yang when he was studying at the university.

China’s ecommerce and cloud computing leader came out with the latest version of its model days after DeepSeek made waves globally, touting better benchmark results. Alibaba, alongside Tencent Holdings Ltd., has a stake in the majority of other contenders on this list even while competing with them. Under Chief Executive Officer Eddie Wu, the company is refocusing on its core strengths, of which cloud computing and the rising AI wave are key components.

China’s most popular consumer chatbot comes from TikTok’s parent company. By its own estimation, ByteDance has some catching up to do in AI and has made it a top priority. Doubao is built using the Mixture of Experts (MoE) machine learning approach, which is common among several AI models in China. Doubao has gained traction in the country thanks to its witty, humanlike conversation and user-friendly interface. Its meeting summaries and image-generation functions have also appealed to young users.

Baichuan’s latest model adopts a technique dubbed deep thinking — a complex reasoning and problem-solving approach that mimics how humans work through problems. It works across text, language and vision. The company offers a narrower evidence-based medical mode, designed to retrieve health-care research and data for medical professionals and academics. Founder Wang Xiaochuan sold his web search engine Sogou to Tencent before starting Baichuan.

Minimax uses what it calls the Lightning Attention mechanism to effectively break down extremely long data sequences into smaller, manageable tiles of data. The model is designed to prioritize the most important parts of a long piece of text, without getting bogged down. Founder Yan Junjie received his Ph.D. from the Chinese Academy of Sciences and went on to further studies at Tsinghua University. He was a vice president at SenseTime before leaving to found Minimax.

Open source like DeepSeek, 01.AI released the Yi-Lightning Mixture of Experts model last October, ranking just behind closed-source models from OpenAI and Google on a closely watched benchmark. The startup, which claims to train its AI models at a much lower cost than industry leaders, is furthest along in commercializing its services and is building solutions for various industry applications. It’s announced a joint development project alongside Alibaba.

Stepfun’s latest product is a trillion-parameter large language model that also relies on the MoE technique. MoE is an approach that divides the model into subsets, each with an expertise in handling specific types of data or tasks, helping improve efficiency by only activating the required expert sub-network per task. CEO Jiang Daxin previously headed up Microsoft Corp.’s research efforts in Asia. A Shanghai government-backed fund was among the first backers of the venture.