Former OpenAI research scientist launches new AI model for Tencent

At a glance:

• Tencent’s Hy3 AI model, developed with OpenAI alum Yao Shunyu, targets US rivals with improved reasoning and coding capabilities
• DeepSeek’s V4 Flash/Pro series introduces Hybrid Attention Architecture for long-conversation memory
• China’s $5B+ AI investment surge fuels competition between Tencent, DeepSeek, and ByteDance

Tencent’s Hy3 Model: A Strategic Upgrade

Tencent’s latest AI model, Hy3, represents a significant leap in its quest to challenge US dominance in large language models (LLMs). The update, announced this week, focuses on enhanced complex reasoning and coding capabilities, positioning it as a direct competitor to models like OpenAI’s GPT-4 and Anthropic’s Claude 3. Yao Shunyu, a former OpenAI research scientist who joined Tencent in 2023, leads the project, bringing expertise in transformer architectures and multi-modal AI systems.

The model’s preview version is already being tested by Chinese developers, with early benchmarks showing a 15% improvement in code generation tasks compared to its predecessor, Hy2. Tencent’s investment in AI has doubled to over $5 billion this year, with a focus on open-source frameworks to bypass US export controls. The company has also restructured its research teams to prioritize high-quality training data, a critical factor in LLM performance.

DeepSeek’s V4 Series: A Competitive Counterstrike

In response to Tencent’s advancements, DeepSeek unveiled its V4 Flash and Pro Series models, building on the success of its R1 AI model launched in January 2025. The V4 series introduces a Hybrid Attention Architecture that improves context retention across multi-turn conversations, a key differentiator in enterprise applications.

DeepSeek’s rapid rise has disrupted China’s AI landscape, with its models gaining traction in sectors like finance and healthcare. The company’s open-source approach has attracted developers seeking alternatives to US-controlled platforms. According to industry analysts, DeepSeek’s V4 models could capture 20% of the Chinese LLM market within 12 months, driven by cost efficiency and performance.

China’s AI Arms Race Intensifies

China’s push for AI self-reliance has accelerated since the US tightened export restrictions on advanced chips in 2023. The government’s $50 billion National AI Development Fund now supports over 100 startups, including Tencent-backed Moonshot AI and StepFun.

The competition extends beyond model development to infrastructure. Tencent’s cloud division, which hosts Hy3, has seen a 40% year-over-year increase in AI workloads. Meanwhile, DeepSeek’s partnership with NVIDIA to optimize inference speeds on local GPUs highlights the race for hardware-software integration.

Open-Source vs. Closed Ecosystems

Tencent’s Hy3 and DeepSeek’s V4 models both embrace open-source principles, contrasting with OpenAI’s closed ecosystem. However, Tencent’s strategy includes proprietary elements, such as its Hunyuan LLM training framework, which combines public datasets with internal corporate data.

This hybrid approach allows Tencent to maintain control over sensitive applications while fostering community contributions. Yao Shunyu emphasized in a recent interview that open collaboration accelerates innovation but warned of risks in data privacy.

Implications for Global AI Development

The rivalry between Tencent and DeepSeek underscores a broader shift in AI geopolitics. As US-China tensions persist, open-source models are becoming battlegrounds for technological sovereignty.

Analysts predict that by 2026, 60% of enterprise AI deployments in China will rely on domestic models like Hy3 and V4. However, challenges remain in matching the scale and polish of US models, particularly in areas like multilingual support and real-time reasoning.

What’s Next for AI Startups

Tencent’s investment in Moonshot AI and StepFun signals confidence in China’s startup ecosystem. Moonshot’s recent $200 million funding round aims to develop specialized AI agents for enterprise workflows, while StepFun focuses on AI-driven content creation tools.

DeepSeek, meanwhile, plans to expand its V4 series to include industry-specific variants for healthcare and legal sectors. The company’s open-source model licenses have already been adopted by over 50 Chinese firms, according to its latest report.

Long-Term Outlook

The race between Tencent, DeepSeek, and US rivals will shape the next decade of AI development. As models grow larger and more specialized, the focus will shift to efficiency, ethical AI practices, and cross-border collaboration.

Yao Shunyu’s move from OpenAI to Tencent exemplifies the fluidity of AI talent in this high-stakes environment. With China’s $50 billion AI fund and US companies racing to maintain dominance, the global AI landscape is poised for transformative change.