DeepSeek’s New AI Model Brings Fight to Alibaba

DeepSeek is said to be developing a new model with advanced AI agent features to compete with US rivals such as OpenAI. Bloomberg Intelligence senior analyst Robert Lea says the planned model should help it catch up to Alibaba, Tencent and ByteDance

DeepSeek: The Rising AI Challenger

• Budget-friendly performance
  DeepSeek’s V3 model, released in December 2024, uses a mixture-of-experts architecture with 671 billion parameters (37 billion activated per token). It was trained using only around 2.8 million GPU hours on H800 GPUs, costing approximately US $5.58 million—a fraction of what models like Meta’s require. Sophisticated architecture with open access
  It employs Multi‑Head Latent Attention (MLA) and a MoE (Mixture-of-Experts) sparse structure, delivering strong performance at low cost. The company also adopted an open-source strategy to broaden research access.

• Global market impact
  The R1 reasoning model, launched in January 2025, reportedly matched or exceeded OpenAI’s o1 in areas like math, science, and coding for far less expense.
  Industry leaders, including Satya Nadella and Sam Altman, called its efficiency “super impressive,” though Altman questioned claims of cost-effectiveness.

• Setback in hardware deployment
  DeepSeek faced delays launching its R2 model due to hardware compatibility issues with Huawei’s Ascend chips. Training had to revert to Nvidia hardware, highlighting challenges in relying on domestic chips.

• Latest upgrade
  DeepSeek has since rolled out DeepSeek‑V3.1, featuring hybrid inference structure, faster processing, enhanced agent capabilities, and updated API pricing effective September 6, 2025.

Alibaba’s Response: Hitting Back Hard

• Qwen 2.5‑Max release
  On January 29, 2025, Alibaba’s cloud arm unveiled Qwen 2.5‑Max, claiming it outperforms DeepSeek‑V3 (as well as GPT‑4o and Meta’s Llama‑3.1‑405B) across major benchmarks.

• QwQ‑32B debut
  In March 2025, Alibaba launched QwQ‑32B, a 32‑billion parameter reasoning model that the company says matches or rivals DeepSeek‑R1, yet runs on far fewer parameters—implying greater efficiency.

• The market responded enthusiastically—Alibaba’s shares jumped around 8% in Hong Kong trading, hitting a 52-week high.

• Broader AI ambitions
  Alibaba is investing over US $50 billion into AI and cloud infrastructure over the next three years, driven by a surge in AI service revenue and the launch of its new AI chip.

What’s Driving This AI Clash?