Large Model Emu3 "Debuts Brightly" 7 Oversold Concept Stocks Attract Financing Investors

Chinese large models are rapidly entering the world’s top tier.

Leading AI chip stocks plummet at the open

Today (February 3), the new energy sector rebounded across the board, with concepts related to solar energy such as BC batteries, TOPCon batteries, HJT batteries, and perovskite batteries leading the gains. BC battery concept stock Aotewei hit the “20cm” daily limit, with Dike Co., Ltd., Roboteck, Laplace, Liancheng CNC, Foster among the top gainers.

AI chip concepts suffered heavy losses at the open. The leading stock Cambricon-U once fell over 14% during trading, with the latest share price dropping below 1100 yuan per share. At yesterday’s close, the company’s A-share market capitalization was 523.733 billion yuan. After today’s sharp decline, it has fallen below 460 billion yuan.

Recently, the company released an earnings forecast indicating that net profit in 2025 is expected to be between 1.85 billion and 2.15 billion yuan, turning from loss to profit year-over-year. Based on the midpoint of the forecast, the company is expected to earn 395 million yuan in Q4 2025, a 30.34% decrease quarter-over-quarter.

Regarding the company’s stock price plunge, a staff member from Cambricon’s secretarial office responded to the media: “We are not very clear about the reason for this (stock price fluctuation). First of all, there are many rumors in the market, and they are definitely false. Then, the fluctuations in secondary market funds and sentiment are still quite significant. Everyone should approach rationally.”

Additionally, two other AI chip leaders, Muxi Co., Ltd. and Moores also fell over 5% at the open, with their lowest prices during the session hitting new lows since listing.

Multi-modal Large Model from Zhiyuan

Emu3 Debuts in Nature

According to Science and Technology Daily, on January 28, the multi-modal large model “Emu3” led by Beijing Zhiyuan Research Institute was officially published in the top-tier international academic journal Nature (the print version is expected to be published on February 12). This marks the first time a large model led by a Chinese research institution has appeared in this journal, signifying a major breakthrough in China’s original AI innovation.

Previously, language large models achieved significant breakthroughs relying on the autoregressive approach of “predicting the next token (NTP).” However, multi-modal models still depend on contrastive learning, diffusion models, and other specialized methods. Whether autoregression can become a universal approach for multi-modal AI remains an industry mystery.

The Zhiyuan team proposed the Emu3 model, which unifies text, images, and videos into the same discrete representation space. It is trained from scratch using a single Transformer architecture, achieving unified multi-modal generation and perception solely through “predicting the next token.”

Experiments show that Emu3 performs on par with diffusion models in text-to-image tasks, and its visual-language understanding rivals the fusion of CLIP and large language models. It can also generate high-fidelity videos using purely autoregressive methods, supporting video extension, text-image interleaving, and robotic operation modeling. Nature’s editors commented that this achievement is of great significance for building scalable, unified multi-modal intelligent systems.

As a core achievement of the “Wujie” large model series, the breakthrough of Emu3 establishes the autoregressive approach as a unified path in generative AI.

Domestic AI Large Models Rapidly Evolving

By 2025, China’s open-source models are emerging as a new force in the global AI landscape. According to Xinhua News Agency, with breakthroughs in technological innovation, application deployment, and ecosystem building, Chinese large models are racing into the top tier worldwide.

A joint report by MIT and the open-source platform Hugging Face shows that over the past year, the global download share of Chinese-developed open-source models reached 17.1%, surpassing the US at 15.8%, ranking first globally.

While gaining international recognition, domestic AI large models continue to upgrade. Recently, several Chinese model companies updated their models almost simultaneously. Alibaba released the Qwen3-Max-Thinking reasoning model, followed by DeepSeek and Kimi updating their models: DeepSeek open-sourced the new DeepSeek-OCR 2, and Kimi released and open-sourced the Kimi K2.5 model.

On February 3, Zhipu announced the official release and open-source of GLM-OCR. The model has only 0.9 billion parameters and supports deployment via vLLM, SGLang, and Ollama. It achieved SOTA performance on multiple benchmarks including formula recognition, table recognition, and information extraction.

From recent updates by major domestic model providers, the focus has shifted from simply comparing parameters and dialogue performance to deeper engineering and system-level capability reshaping.

13 AI Large Model Concept Stocks

Performance Expectations

Based on disclosed earnings forecasts and performance reports, 13 AI large model concept stocks are expected to see net profit growth in 2025 (including those turning from loss to profit), with Jadu Technology, Yili Media, Lio Co., Ltd., Santai Co., Ltd., Aofei Entertainment among the top performers in profit growth.

Jadu Technology expects net profit for 2025 to be between 160 million and 200 million yuan, turning profitable from loss. During the reporting period, the company continued to strengthen its AI technology in intelligent transportation, maintaining stable business expansion and order delivery. Additionally, the fair value change of its AI ecosystem listed companies impacted net profit by 63 million yuan (after tax).

Yili Media expects net profit for 2025 to be between 14 million and 20 million yuan, turning profitable. During the period, benefiting from strong demand from internet social industry clients, daily chemical and beauty industry clients, and 3C industry clients, the company achieved high growth despite challenging consumer environments. It also launched its self-developed AIGC creative platform “Jiuhé Creativity” and promoted the deep application of AI tools in content production, strategic planning, and operations, driving cost reduction and efficiency.

Lio Co., Ltd. expects net profit for 2025 to be between 190 million and 250 million yuan, turning profitable. The company recently stated in an institutional survey that it has AI application capabilities for multi-modal content in relevant businesses, including not only text but also images, charts, and videos.

On the capital side, several oversold AI large model stocks have seen increased holdings from institutional investors.

According to Data Treasure, as of February 2 close, 22 AI large model concept stocks had net financing net purchases exceeding 100 million yuan since January. Compared with their year-to-date highs, seven stocks have retraced over 20%, including Chinese Online, Tuors, Kunlun Wanyou, Tom Cat, Digital Tech, Chuangye Huikang, Hande Information.

(Source: Data Treasure)