Chinese artificial intelligence companies are rapidly releasing model updates ahead of the Lunar New Year holiday, showcasing significant advancements led by the start-up DeepSeek amidst US chip restrictions.
On the eve of China’s key annual festival, DeepSeek launched a new open-source model for image generation, solidifying its role as a major innovator in an area historically dominated by US firms. This new release follows similar updates from Alibaba and start-ups Moonshot and Zhipu.
One product manager at a language model start-up remarked that the push for releases before the holiday was akin to making a big announcement on Christmas Eve, highlighting the urgency in the sector.
DeepSeek’s progress has elicited concern in the US about the rapid advancements of Chinese laboratories operating with limited budgets, while also fostering an environment of newfound confidence within the Chinese AI landscape that could lead to increased investments. An AI investor noted that DeepSeek has surpassed other Chinese model companies in terms of progress, igniting belief that they can close the gap with global leaders.
DeepSeek has gained international recognition with a series of model releases that demonstrate performance comparable to top US competitors like OpenAI and Meta, despite claiming to operate with fewer resources and being hindered from accessing the latest Nvidia chips due to US export limitations. Recently, the company introduced its R1 reasoning model, which can autonomously learn and improve without human intervention.
According to an engineer at AI research hub Hugging Face, DeepSeek’s developments are revitalizing Chinese AI initiatives and contributing positively to the global open-source AI community, as its findings are expected to enhance reasoning models.
In the wake of these breakthroughs, investors reacted by selling AI-related stocks, with Nvidia experiencing a loss of nearly $600 billion in market value. This response reflects concerns over the feasibility of producing robust models in a manner distinct from the US strategy focused on scaling up computing power.
On the same day as DeepSeek’s new release, Alibaba’s Qwen introduced Qwen2.5-1M, featuring capabilities to handle longer inputs, which is vital for AI applications requiring substantial memory. Additionally, DeepSeek unveiled Janus-Pro, a text-to-image generation model reportedly surpassing competitors like OpenAI’s Dall-E 3 on certain benchmarks.
Zhipu, which was valued at $3 billion in December, also updated its GLM-PC model aimed at enterprise applications, enabling more efficient automated task completion.
Meanwhile, another Beijing-based start-up, Moonshot, updated its Kimi reasoning model to Kimi k1.5, showing promising results in complex reasoning tasks and supporting both text and image processing.
Chinese tech firms typically launch products before lengthy holidays, allowing potential customers ample time to engage with new offerings. Following the holiday, competition will intensify among Chinese AI players to dominate the development of commercially viable AI applications, as some believe that AI agents could yield substantial commercial value, potentially transforming one or two language model companies into leaders in next-generation software development.
photo credit: www.ft.com
