The AI world was ignited once again by Alibaba on Lunar New Year's Eve. On February 16, the company open-sourced its new-generation large language model Qwen3.5-Plus, which boasts performance comparable to top-tier closed-source models like Gemini-3-pro and GPT-5.2, marking the beginning of a competitive "model supremacy" spring offensive.
In the early hours of the first day of the Lunar New Year, February 17, data released by Qwen revealed that during the Spring Festival activities, over 130 million people in China experienced AI-powered shopping for the first time, uttering the phrase "Qwen, help me" five billion times. This surge has elevated Qwen to the status of a national-level AI assistant.
Unlike the industry's common practice over the past year of stacking parameters and scaling up model size, Qwen3.5 represents Alibaba's breakthrough strategy of achieving superior performance with a more compact architecture. It is understood that Qwen3.5 has undergone a comprehensive overhaul of its underlying model architecture. The Qwen3.5-Plus version features a total of 397 billion parameters but activates only 17 billion, outperforming the trillion-parameter Qwen3-Max model. This design reduces GPU memory usage for deployment by 60% and significantly improves inference efficiency, with maximum inference throughput increasing by up to 19 times. Notably, Qwen3.5-Plus's API pricing is as low as ¥0.8 per million tokens, merely 1/18th the cost of Gemini 3 Pro.
Behind the technological advancements of Qwen3.5 and the rapid user growth of the Qwen app, Alibaba has made another critical move in the battle for dominance over the infrastructure of the AI era.
The arrival of Qwen3.5 is more than a simple model iteration; its release during the Spring Festival signifies a new window of opportunity for the industry as the benefits of Scaling Law begin to plateau. Unlike previous generations of Qwen large language models, Qwen3.5 represents a generational leap from a pure text model to a natively multimodal model. It is pre-trained on a mixture of visual and text tokens and has seen substantial additions of data in Chinese, English, multilingual content, STEM, and reasoning. With less than 40% of the parameters, it achieves top-tier performance comparable to the trillion-parameter Qwen3-Max base model. It demonstrates excellent results across comprehensive benchmark evaluations in reasoning, programming, and Agent capabilities, surpassing Gemini 3 Pro and GPT-5.2 in benchmarks like the general Agent evaluation BFCL-V4 and the search Agent evaluation Browsecomp.
Native multimodal training has also led to a significant leap in Qwen3.5's visual capabilities. It has achieved best-in-class performance on numerous authoritative evaluations, including multimodal reasoning (MathVison), general visual question answering VQA (RealWorldQA), text recognition and document understanding (CC_OCR), spatial intelligence (RefCOCO-avg), and video understanding (MLVU). Furthermore, Qwen3.5 shows marked improvements in academic problem-solving, spatial reasoning, and video comprehension, supporting video inputs of up to 2 hours in length. It also achieves native integration of vision and code, enabling capabilities like converting sketches directly into front-end code and using screenshots to locate and fix UI issues, thereby turning visual programming into a practical productivity tool.
The key to Qwen3.5's performance leap lies in major innovative breakthroughs to the classic Transformer architecture. Currently, most existing multimodal models in the industry still rely on stacking modalities and parameters—adding visual and then audio components onto a text model—using scale to compensate for architectural limitations. This approach often leads to model bloat, high inference costs, limited intelligence, and low inference efficiency, making real-world deployment challenging.
For Qwen3.5, Alibaba not only employed a hybrid MoE (Mixture of Experts) architecture but also adopted a novel training paradigm. During pre-training, a unified architecture processes multiple modalities of data simultaneously. With a total of 397 billion parameters and only 17 billion activated, it is the smallest model by parameter count to achieve such performance levels. Additionally, it is understood that the Qwen team's self-developed gating technology, which won the Best Paper award at the top global AI conference NeurIPS 2025, has been integrated into Qwen3.5's innovative hybrid architecture, playing a crucial role in breaking the model's performance ceiling. Through optimizations in training stability and techniques like multi-token prediction, Qwen3.5 matches the performance of the Qwen3-Max model while further enhancing inference efficiency.
Behind this "smaller model, bigger performance" strategy, Alibaba, as a major model developer, is pioneering the transformation of "performance gains" into "scalable, replicable engineering capabilities." This not only improves the model's cost-effectiveness but also furthers its evolution into a genuine productivity tool.
From a broader industry perspective, Qwen3.5 is just the tip of the iceberg in Alibaba's multimodal strategy. The company continues to explore an "All in One" omnimodal approach, pushing towards AGI (Artificial General Intelligence). A week prior, Qwen released Qwen-image-2.0, which integrates image generation and editing capabilities and is seen as a domestic counterpart to Nano Banana Pro. With Qwen3.5's latest breakthrough in visual understanding, Qwen has taken a significant step closer to full omnimodal capability. The parallel advancement of these two technical paths reveals Alibaba's ambition to build momentum for omnimodal exploration. This signifies Alibaba's aim to enable large models to progress from solving single tasks to possessing comprehensive abilities such as "understanding images, comprehending videos, and coordinating tools to execute tasks"—making AI not just capable of seeing and hearing but also of taking action.
In January, the Qwen App launched a consumer-grade AI shopping Agent. During the Spring Festival holiday, this AI shopping Agent helped users complete 120 million orders in just six days. The Qwen App and PC client have now been promptly updated with the Qwen3.5-Plus model. With significantly enhanced Agent capabilities, Qwen3.5 will further expand the app's ability to assist users with tasks in both work and daily life.
Data released by Qwen on February 17 shows that over the past two days, user orders for purchasing tickets via AI increased 22-fold compared to the previous period. Orders for transportation tickets like flight bookings via AI grew more than 7 times. With the release of new movies for the Spring Festival season, orders for buying movie tickets through AI surged 372-fold, with orders from third- and fourth-tier cities seeing a staggering 782-fold increase. Nearly half of all AI orders originated from county-level areas. Due to the convenience of voice-activated ordering, nearly 4 million users aged 60 and above experienced AI shopping for the first time.
Progressing on the path towards AGI, this is the true potential unlocked by Qwen3.5 during the Spring Festival period.
Powered by Alibaba Cloud Behind Alibaba's flurry of activity during the Spring Festival, if Qwen3.5 is likened to the sharp spearhead of its charge towards AGI, then the "powered by Alibaba Cloud" ecosystem represents another, more concealed yet decisive front: the closed-loop ecosystem of AI infrastructure.
While continuously releasing state-of-the-art (SOTA) models, Alibaba has also open-sourced over 400 Qwen models since 2023, covering all sizes and modalities. The global download count for Qwen has exceeded 1 billion, and developers have created over 200,000 derivative models based on Qwen. From teams like Fei-Fei Li's to companies like Airbnb, global developers are building their own AI applications on top of Qwen.
To meet the needs of AI developers and enterprises in different countries, the Qwen large model continues to evolve. Qwen3.5 expands support to 201 languages, increasing its vocabulary size from 150,000 to 250,000 tokens, which can boost coding efficiency for low-resource languages by up to 60%. The prosperity of this open-source ecosystem not only provides Qwen with vast amounts of feedback data but also plants the brand perception of "Qwen as infrastructure" in the minds of developers.
Simultaneously, Alibaba's long-term investments in the infrastructure layer are beginning to create synergistic effects with its model capabilities. On January 29, Alibaba's T-Head officially launched its self-developed AI chip, the "Zhenwu 810E." The Zhenwu PPU utilizes a self-developed parallel computing architecture and inter-chip interconnect technology, coupled with a full-stack self-developed software suite, achieving complete in-house development from hardware to software. It features 96GB of HBM2e memory and an inter-chip bandwidth of 700 GB/s, making it applicable for AI training, AI inference, and autonomous driving. Comparing key parameters, the overall performance of the Zhenwu PPU surpasses NVIDIA's A800 and mainstream domestic GPUs, being comparable to NVIDIA's H20.
The Zhenwu PPU is already being used extensively within Alibaba for training and inference of the Qwen large models. It is understood that the chip has been heavily optimized for mainstream MoE architecture models like Qwen3, meeting the massive computational demands of these large models.
The further reduction in Qwen3.5's cost is also得益于 (benefiting from) synergistic innovation between the model, chip, and cloud. This indicates that the "1+1+1>3" synergistic effect between Alibaba's models, chips, and cloud platform is already materializing in terms of real deployment costs and throughput efficiency. This synergy constitutes Alibaba Cloud's moat in the AI era infrastructure landscape.
For Alibaba Cloud, this also translates into a more stable long-term revenue structure. The "powered by Alibaba Cloud" infrastructure system is building a more sustainable channel for Alibaba within the AI industry, paving the way for commercial success.
Market research firm Omdia reported that in the first half of 2025, the overall Chinese AI cloud market reached ¥22.3 billion, with Alibaba Cloud holding a 35.8% share—exceeding the combined share of the second to fourth players. In Q3 2025, Alibaba Cloud's market share in China's cloud market increased to 36% from 34% the previous quarter, widening its lead.
Omdia noted that AI is increasingly becoming the primary driver of new demand for core cloud infrastructure services, leading to growth in the consumption of computing, storage, and database resources. Alibaba's financial reports show that revenue from Alibaba Cloud's AI-related products has achieved triple-digit year-on-year growth for nine consecutive quarters.
In September 2025, a report by international market research firm Frost & Sullivan indicated that in a survey of 700 enterprises across key industries such as finance, manufacturing, internet, consumer electronics, and automotive, Alibaba's Qwen held a 17.7% share in China's enterprise large model invocation market, ranking first.
From the technological debut of Qwen3.5 marking the Spring offensive, to the synergistic advancement of models, chips, and cloud infrastructure, and further to the generational leap in the Qwen app's Agent capabilities and the consumer phenomenon of 130 million people embracing AI shopping, Alibaba's long-term strategic layout for the AI era is gradually unfolding. On one front lies the "vast cosmos" of pursuing AGI through top-tier models and underlying architectures, constantly pushing the boundaries of intelligence. On the other front is the Qwen App bringing AI services into the daily lives of ordinary people, translating technological benefits into the routines of hundreds of millions of users. Through this Spring Festival "two-front operation," Alibaba is simultaneously supporting the most cutting-edge imagination and grounding it in the most tangible aspects of everyday life.