Chinese artificial intelligence startup DeepSeek is reportedly looking to release a new AI model designed to perform advanced agent tasks better than OpenAI’s ChatGPT, according to people familiar with the matter.
According to a Thursday Bloomberg insight, the Hangzhou-based company is working on software capable of carrying out multi-step actions on behalf of users with “minimal human input.”
The planned system is learning from past shortcomings and could improve its performance over time, the people requesting anonymity told Bloomberg. DeepSeek founder Liang Wenfeng is reportedly pushing his team to release the new agent-focused model in the final quarter of this year.
If launched within the proposed timeframe, it could become the company’s most ambitious move since January’s R1 release, a large language model (LLM) that mimicked human reasoning and was supposedly developed at just $6 million, a small fraction of the cost of ChatGPT’s systems.
DeepSeek is building LLMs on previous releases
R1’s launch may have generated international attention, topping US download charts during its first week; DeepSeek has not yet released a new AI model and only made incremental updates. Rivals in both the United States and China have rolled out a series of new models in the meantime.
As reported by Cryptopolitan on August 21, the company released its flagship V3 system, which it said was optimized for compatibility with Chinese-made chips. According to developers, the software has faster processing speeds and could reduce Beijing’s dependence on American-made chips.
China has been technologically held hostage with export restrictions from Washington, which prevented Nvidia from sending advanced chips to the second-largest economy in the world.
The August launch followed two other updates earlier in the year: an R1 upgrade in May and an enhancement of the V3 platform in March. DeepSeek reiterated that its V3.1 model supports a precision format known as UE8M0 FP8, optimized for “soon-to-be-released next-generation domestic chips.” FP8, or 8-bit floating point, is a data format that allows AI systems to use less memory while increasing speed.
DeepSeek changes terms of use
The company is trying to compete with American AI players such as OpenAI, Anthropic, and Microsoft, who have all introduced their own agent-style platforms in recent months to make artificial intelligence more autonomous and task-driven.
But back home, DeepSeek is also keeping watch of companies like Manus AI, a startup with Chinese roots that has created what it describes as a “general AI agent.”
Unlike traditional chatbots, which generate text-based responses, Manus AI agent systems execute complex tasks like planning travel, writing and debugging software.
Alongside its technical development, DeepSeek has implemented stricter policies governing AI-generated material on its platforms. According to local news outlet GizmoChina, the AI startup announced new requirements mandating that all content created with its models be clearly labeled as artificial.
An official notice shared through the media said these markers include both visible and hidden signs. Visible identifiers could be text such as “AI-generated,” audio disclaimers, or on-screen graphics. Hidden markers are embedded within metadata and may include information about the content type, the company that produced it, and a unique identification number.
DeepSeek has prohibited users from altering, removing, or falsifying these labels. It has also banned tools designed to tamper with the markings, warning that violators may face legal consequences, backed by the Chinese government’s laws.
On the profit side, DeepSeek’s rout at the start of the year improved China’s AI market, and Huawei Technologies Co., the Shenzhen-based telecommunications and electronics giant, credited the AI demand for helping it return to profitability in the first half of the year.
