DeepSeek R2 is coming fast: Can the West keep up?

DeepSeek is set to accelerate the launch of its new AI model, R2, following the success of its previous model, R1, which recently prompted a $1 trillion sell-off in global equities markets due to its competitive performance against Western counterparts.

DeepSeek accelerates launch of R2 AI model after R1 success

The Hangzhou-based startup originally aimed to release R2 in early May but is now pushing for an expedited rollout, according to three sources familiar with the company. The new model is expected to enhance coding capabilities and support reasoning in multiple languages beyond English. DeepSeek has not commented on these developments.

R1, developed using less powerful Nvidia chips, remains competitive with models created by major U.S. tech firms that have invested hundreds of billions in AI technologies. Vijayasimha Alilughatta, COO of Zensar, stated that the release of R2 could significantly impact the AI industry, challenging the dominance of existing market leaders.

Founder and operational ethos

Liang Wenfeng, founder of DeepSeek and a billionaire from his quantitative hedge fund High-Flyer, has kept a low profile since July 2024. Former employees describe DeepSeek as operating more like a research lab than a conventional business, characterized by a flat management style that contrasts with the hierarchical structure common in China’s tech sector.

Liang, who began his career in smart imaging and later managed a research team, was praised for hiring top algorithm engineers and fostering a collaborative environment. Employees noted that many worked eight-hour days, feeling empowered to contribute to key projects. Liang emphasizes quality over extensive app development, having commented on avoiding heavy spending in favor of refining AI models.

Compensation at DeepSeek and High-Flyer is reportedly generous; senior data scientists at High-Flyer can earn up to 1.5 million yuan annually, significantly higher than competitors. High-Flyer itself is recognized as one of China’s top quant funds, managing tens of billions of yuan despite regulatory pressures on the sector.

DeepSeek rocked the market once and here’s why it could happen again

DeepSeek’s competitive edge stems from High-Flyer’s significant investments in research and computing power. The firm allocated 70% of its revenue towards AI research, building two supercomputing AI clusters, including one consisting of 10,000 Nvidia A100 chips during 2020 and 2021. The Chinese government initially raised concerns about the massive chip acquisition but ultimately allowed operations to proceed, which later benefited DeepSeek when the U.S. banned A100 chip exports to China in 2022.

With limited competition for such resources, DeepSeek has attracted leading researchers. The application of cost-effective techniques, such as Mixture-of-Experts (MoE) and multihead latent attention (MLA), has enabled the company to lower computing costs while remaining competitive. DeepSeek’s pricing is estimated to be 20 to 40 times cheaper than similar models from OpenAI, according to Bernstein analysts.

Since the introduction of R1, Western tech giants have adjusted their strategies, reflected in price cuts from OpenAI and the introduction of discounted access tiers by Google. Benchmark testing from UST noted that R1 often required three times the tokens for reasoning compared to OpenAI’s scaled-down models.

Government support

DeepSeek has gained favor with the Chinese government, as evidenced by Liang attending a high-profile meeting with Premier Li Qiang as a representative of the AI sector. Local governments and several state-owned enterprises have begun integrating DeepSeek’s models, indicating a strong endorsement from Chinese leadership, which aims to demonstrate superiority over U.S. technology.

This rise comes at a time when DeepSeek’s models face removal from app stores in some countries, including South Korea and Italy, due to privacy concerns. Experts warn that if DeepSeek becomes the preferred AI model for Chinese state entities, Western regulators may respond with heightened restrictions on AI technology exports and collaborations.

Liang has acknowledged the challenges posed by the embargo on advanced AI chips, emphasizing funding has not been the primary issue for DeepSeek.

Featured image credit: Matheus Bertelli/Pexels