Introduction
As 2026 passes its halfway mark, the competition within China’s large language model (LLM) sector has entered a new phase. The focus of public attention has shifted noticeably. In mid-July, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that became the world’s largest open-source LLM at the time. Around the same time, DeepSeek rolled out the full-capacity V4 model for internal testing, with benchmark performance matching GPT-5.6.
Data from QuestMobile shows that Doubao, developed by ByteDance, hit 382 million monthly active users (MAUs) in June, securing the leading position among consumer-facing AI products. Meanwhile, Qianwen from Alibaba and Yuanbao from Tencent, two headline-making players in 2025, saw a sharp drop in media exposure this year.
Compared with the market landscape of 2023, the 2026 AI competition has moved past the early exploratory stage. Leading companies now launch iterative upgrades one after another to capture public interest. Moonshot’s Kimi K3 set a new industry bar: it is the first open-source model with over 2 trillion parameters, and its evaluation scores outperform both GPT-5.6 and Anthropic’s Fable 5, raising the entry threshold for open-source competition. DeepSeek’s V4 series delivered substantial performance gains from preview versions to full releases, and introduced cost-effective billing plans, attracting large numbers of small teams and independent developers. ByteDance’s Doubao, supported by its massive existing user base, has grown far beyond the scope of a productivity tool, occupying a large share of fragmented daily user scenarios and dominating online discussion.
This article analyzes the shifting traffic pattern, different development strategies of leading Chinese LLM products, and the internal structural adjustments within Alibaba and Tencent that have led to reduced public exposure for Qianwen and Yuanbao. It also explores the logic behind the second half of the “Hundred-Models War” and the new competition logic between startup teams and tech giants.
Reduced Exposure Does Not Equal Stagnation
Judging purely by media buzz, it is easy to conclude that Qianwen and Yuanbao have fallen behind their competitors. However, operational data tells a very different story.
Alibaba’s Qianwen recorded 167 million MAUs in June, a year-on-year increase of 5792.9%, which pushed it up to the second place in the consumer AI product ranking. The growth momentum continued into April, when the product achieved “recovery growth” after a period of adjustment.
On the product iteration front, Alibaba rolled out a preview version of the new Qianwen model in July, built with the third-generation Mixture of Experts (MoE) architecture featuring 2.4 trillion parameters. The team also launched a subscription plan with discounted computing power to expand commercial adoption. On the B2B side, China FAW built an in-vehicle AI assistant based on Qianwen within Alibaba’s cloud computing platform. The product was also embedded into hardware devices developed by the smartphone manufacturer Yunqi.
Tencent’s Yuanbao also kept a steady iteration rhythm. On July 6, Yuanbao integrated the Hunyuan Hybrid 3 model and unlocked free Agent functions for all users. Then on July 15, it opened up cross-platform interaction with JD’s AI Agent through mini-programs, successfully turning pure chatbot products into transaction-oriented entry points for e-commerce business.
The cooling public perception of Qianwen and Yuanbao stems from the overall market trend. The “Hundred-Models War” has entered its second half. The industry is undergoing a reshuffle: a large number of small and medium-sized vendors have exited the competition, and capital as well as media attention have become more concentrated. The overall “noise” in the industry has faded, so even major vendors receive less public coverage.
Internal organizational changes have also slowed the pace of new product releases. Alibaba restructured its product team after integrating multiple business lines into a unified Qianwen brand. Tencent adjusted the business division responsible for Yuanbao, redefining the product’s positioning within the company’s ecosystem. Both teams have focused on internal coordination, and the output of visible public achievements has temporarily slowed down during the adjustment period.
The Second Half of AI Competition: Business Ecosystems Replace Public Hype
The core competition logic has undergone a fundamental shift as the AI race enters its second phase.
In the first half of the “Hundred-Models War”, startup companies set the agenda for the whole industry. They kept launching new technical labels and larger parameter models to quickly open market gaps. For large tech groups like Alibaba and Tencent, it was impossible to focus all resources on pure model iteration due to the constraints of existing business chains. As a result, they often fell behind startups in viral public marketing campaigns.
In the second half of the competition, the power to set topics gradually shifted to established tech giants.
- ByteDance’s Doubao: It leverages the platform’s existing traffic and cross-scenario synergy to continuously expand user activity.
- Alibaba’s Qianwen: It relies on Alibaba Cloud’s infrastructure and closed-loop e-commerce scenarios to capture B-end enterprise customers steadily.
- Tencent’s Yuanbao: It connects WeChat mini-programs and social traffic to build a complete link leading to real-world transactions.
The whole industry has shifted its focus from technical parameters to three core dimensions: end-user scenarios, enterprise commercialization, and stable ecosystem operation. The inherent resource advantages of big tech firms are turning into long-term competitive barriers. The relatively low profile of Qianwen and Yuanbao is actually a proactive strategic choice: instead of chasing viral hot topics, the two teams prioritize polishing business landing results.
Startups including Moonshot AI and DeepSeek still maintain their pace of technical innovation, but they also face tighter profit pressure. Without mature business scenarios, it becomes harder for pure model providers to secure continuous financing amid the cooling capital market. Many small and mid-sized teams have started to provide customized model services for vertical industries instead of releasing general-purpose large models to compete with giants.
Teams managing multi-vendor model access often need to unify calling standards across dozens of LLMs. An API gateway like 4sapi can simplify the access process, helping developers switch between GPT, Claude, DeepSeek and domestic models without repeatedly modifying code. This kind of unified scheduling capability lowers the trial cost for both enterprise clients and small developer teams.
New Market Pattern: Three Giants and Emerging Startups
After two years of fierce competition, China’s LLM market has formed a clearer tiered structure. The first tier is occupied by three large tech groups: ByteDance, Alibaba and Tencent. Their products cover both C-end consumer traffic and B-end enterprise services, with complete cloud infrastructure and scenario support. The second tier consists of emerging startup leaders, represented by Moonshot AI and DeepSeek. They maintain leading advantages in open-source model technology and developer ecosystems, and have built stable paid user groups among technical teams. The third tier includes vertical industry models and lightweight fine-tuned products. These teams give up the general-purpose model track and focus on industry-specific scenarios such as finance, manufacturing and education.
Doubao has secured the top position in the C-end market with its traffic advantage. Its monthly active user base continues to expand, far outpacing most competing consumer AI products. Qianwen and Yuanbao, after short-term team restructuring, have stabilized their growth curves and started to dig deep into the synergy between models and their parent companies’ core businesses. Qianwen tightly links with Alibaba’s cloud and e-commerce business, while Yuanbao keeps exploring the commercial potential of social and mini-program ecosystems.
For open-source track players, Kimi K3 and DeepSeek V4 have become two core technical benchmarks. The continuous iteration of these two models keeps lowering the threshold for downstream developers to build secondary applications, which in turn expands the whole developer ecosystem. As more teams build applications based on these open model weights, the moat for Moonshot and DeepSeek keeps widening.
Conclusion
The shifting spotlight in the 2026 AI competition reflects a natural evolutionary phase of the LLM industry. The era of simply competing on parameter size and technical buzzwords is coming to an end.
ByteDance’s Doubao takes the lead in consumer traffic, while Moonshot and DeepSeek lead the open-source technical competition. Alibaba’s Qianwen and Tencent’s Yuanbao temporarily step out of the public limelight, focusing on internal team adjustment and business landing. Their reduced media exposure is not a sign of decline, but a strategic turn toward sustainable commercial growth.
In the second half of the Hundred-Models War, the core competition will center on scenario implementation, commercial monetization and ecosystem construction. Tech giants with complete industrial chains will gain more advantages, while outstanding startups can still break out by taking root in open-source communities and vertical niche markets. The pattern of “three giants plus innovative startups” will remain stable for a long time, and the next round of reshuffling will happen among teams that can turn model capability into real business value.
