Goldman’s First-Take On Alibaba’s Hybrid Qwen3 Model

The artificial intelligence race to outperform Chinese DeepSeek intensified on Tuesday as Alibaba unveiled Qwen3, a family of open-source large language models. Goldman analysts told clients the new LLMs represent a continued boom in China’s AI space.

On Tuesday, the Qwen team published a blog post on its website announcing the results of Qwen3-235B-A22B, showing that the flagship model achieved competitive results in benchmark evaluations of coding, math, general capabilities, etc., when compared to other LLMs, including DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro. 

According to Qwen, Qwen3-235B-A22B transitions between a “thinking mode” and a “non-thinking mode” in a hybrid approach to problem-solving. 

I seriously cannot believe this is a 0.6B LLM! @Alibaba_Qwen just released Qwen3, a series of hybrid reasoning models that allow you to control how much “thinking” the model does for a given task.

They can even run locally in your browser on WebGPU with Transformers.js! pic.twitter.com/jnjWanObw9

— Xenova (@xenovacom) April 29, 2025

Goldman analysts Ronald Keung, Timothy Zhao, and colleagues commented on the rise of China’s AI models like Qwen, outlining seven key observations for clients after reviewing the Qwen3-235B-A22B:

1. Its leading benchmarking results for its flagship model (Qwen3-235B-A22B) in coding, math etc. vs. other top-tier models DeepSeek-R1, o1, etc.,

2. further improved performance alongside lower inference cost, with the smaller MoE model (Qwen3-30B-A3B) offering higher performance with much less activated parameters,

3. its smaller models allow for flexibility of local deployment and edge applications across mobile devices, smart glasses, autonomus vehicles, robotics etc.,

4. integrated thinking (complex reasoning) and non-thinking (instant responses) modes, offering users/enterprises to dynamically manage the inference costs.

5. expanded pre-training dataset for Qwen3 (36tn token, vs. 18tn token of Qwen2.5),

6. more AI agentic capabilities with support of MCP (model context protocol), and

7. global accessibility, supporting 119 languages and dialects. Separately, Alibaba Cloud announced earlier at its 2025 AI conference on April 9 where API calls for AI models on Alibaba Cloud surged nearly 100X in Jan 2025 vs. Dec 2024, while the number of enterprises integrated with the PAI model platform has also grown 100X to over 10k vs. last year

The analysts are “Buy” rated on BABA with a 12-month price target of $159 “on stabilizing domestic eCommerce profits and fast Alibaba Cloud revenue growth being China’s largest cloud hyperscaler, with leading AI models and diverse application scenarios.”

The race for more efficient LLMs will likely lead Washington to continue tightening chip exports to curb China’s advances, especially with developments from DeepSeek and Alibaba. US exports of Nvidia’s H20 AI chips were recently banned, while US-blacklisted Chinese Huawei Technologies said Monday that it’s testing its new powerful AI chips that exceed the performance of H100s. 

“Alibaba’s release of the Qwen 3 series further underscores the strong capabilities of Chinese labs to develop highly competitive, innovative, and open-source models, despite mounting pressure from tightened U.S. export controls,” Washington-based analyst Ray Wang told CNBC. 

Wang warned: ”The U.S.-China AI race, the gap between American and Chinese labs has narrowed—likely to a few months, and some might argue, even to just weeks.” 

“With the latest release of Qwen 3 and the upcoming launch of DeepSeek’s R2, this gap is unlikely to widen—and may even continue to shrink,” he noted.

In markets, the Nasdaq 100 began spiraling lower just a week or so after DeepSeek’s R1 was launched in late January. 

What happens to the Nasdaq on a DeepSeek’s R2 launch? 

Well, for one thing, we know peak data center capacity has likely arrived:

• Wells Fargo Analysts Say Amazon Paused Some Data Center Lease Commitments

• Goldman Throws Cold Water On AI Hype, Moves Forward Datacenter Peak Forecast

. . .