Three Chinese AI Giants, Three Strategies: Doubao, DeepSeek, and Qwen in 2025

In 2025 China's AI large‑model battlefield is undergoing deep segmentation. Using user activity as a metric, ByteDance’s Doubao leads in engagement; DeepSeek remains the technical and international leader; Alibaba’s Tongyi Qwen shows a unique strategic advantage in the full‑stack layout.

Doubao: Capturing Users with "Everyday Magic"

Launched in August 2023 with a low‑key rollout, Doubao leverages ByteDance’s super‑apps (Douyin, Toutiao) as its backyard, embedding AI into short‑video, recommendation and customer‑service scenarios, making AI as accessible as utilities. QuestMobile September data shows Doubao, DeepSeek and Tencent Yuanbao with 172 M, 145 M and 32.86 M monthly active users respectively, with Doubao’s rapid growth driven by internal traffic.

Technically, Doubao does not chase parameter bloat; it adopts a MoE + UltraMem sparse architecture that maximizes speed and cost efficiency, delivering instant responses and low billing. Its "instant‑understanding" of ordinary users makes it simple, fun and barrier‑free, turning AI into a "social currency" for Gen‑Z—photo editing, voice imitation, chat companionship, etc. On 1 December 2023 Doubao partnered with ZTE to launch the nubia M153 (¥3499), embedding AI in a pocket device and redefining "ground‑level flight".

However, the "everyday magic" is a double‑edged sword: users become accustomed to free usage, making paid conversion difficult, and the model’s presence remains confined to content‑entertainment scenarios, struggling to penetrate high‑value domains such as finance, healthcare and education. The next hard battle for Doubao is to translate C‑side popularity into B‑side revenue streams.

DeepSeek: Open‑Source Vanguard Claiming AI Sovereignty

At the start of 2025 DeepSeek announced the official launch of V3, fully open‑source. Benchmarks show DeepSeek‑V3 surpasses Qwen2.5‑72B and Llama‑3.1‑405B, and its performance rivals GPT‑4o and Claude‑3.5‑Sonnet, breaking Western monopoly in AI technology.

On 2 December 2025 DeepSeek released V3.2 and V3.2‑Speciale; V3.2 already matches GPT‑5 in public evaluations, while the high‑end Speciale version approaches Google Gemini 3 Pro in inference scores.

Training data includes 24 667 real code tasks, 50 275 real search logs and thousands of synthetic scenarios, enabling the model to embed a "think‑step‑tool‑then‑advance" workflow and boost agent capabilities. The release package contains model weights, de‑identified training data and the inference framework, causing GitHub stars to soar and cementing DeepSeek as one of the hottest Chinese base models.

The open‑source strategy follows three moves: (1) instantly lower entry barriers so developers can build on DeepSeek without retraining large models; (2) attract national‑level products (Baidu, Alibaba, NetEase) to adopt the model, creating a healthy AI food chain; (3) use openness to challenge Western tech walls and broadcast China’s ability to produce top‑tier models while sharing the "cake" globally.

Qwen (千问): Full‑Stack Play Betting on the Intelligent Future

Alibaba’s AI rollout lags behind Doubao and DeepSeek in speed, but its mainstream C‑side products only scratch the surface, leaving deep pain points and fragmented scenarios. Qwen therefore embeds itself completely into Alibaba’s economic ecosystem, aiming to be an AI that can both chat and get things done.

On 17 November 2025 the Qwen app entered public beta, featuring the "world’s first open‑source model" Qwen 3 and directly challenging ChatGPT. Within a week the beta surpassed ten million downloads, outpacing the early growth of both ChatGPT and DeepSeek.

Qwen 3‑Max has overtaken GPT‑5 and Claude Opus 4 in multiple benchmarks, while the Qwen series overall has accumulated over 600 million global downloads since its 2023 open‑source release, surpassing Llama and DeepSeek in popularity. Backed by a ¥380 billion AI infrastructure plan announced in February 2025, Alibaba intends to expand cloud compute capacity tenfold by 2032, essentially buying compute freedom with massive spending.

The real strength of Qwen lies in its integration across Alibaba’s business matrix: Taobao and Tmall for shopping, DingTalk for work, Fliggy for travel, Amap for navigation, etc. When competitors are still hunting for scenarios, Qwen has already woven AI into payment, shopping, ticketing and navigation at a granular level, creating a seamless, "invisible" user experience. Future examples include ordering the fastest‑delivered lunch on Taobao with a single voice command, or planning a weekend trip to Hainan via Fliggy, or getting optimal navigation routes that avoid highways via Amap.

Conclusion: Beyond Parameters, the Real Competition Is Scenario, Ecosystem, and Monetization

In the AI race of 2025, the winner will not simply be the model with the most parameters or the fastest response, but the one that masters user experience, locks in traffic, and builds a sustainable revenue engine. Doubao leverages ByteDance’s product buffs to dominate C‑side engagement; DeepSeek uses full open‑source to win technical sovereignty and the developer community; Qwen exploits Alibaba’s massive ecosystem to create a full‑stack AI layer that becomes part of daily life. Emerging paradigms such as embodied intelligence, multimodal reasoning and AI agents will further reshuffle the field, but China’s AI sector has already moved from "usable" to "delightful", pushing the industry toward a future where AI is as essential as air and water.