Rival Realities: AIs New Year Clash

China's AI Giants Showcase Divergent Paths in Fierce Lunar New Year Showdown

The Lunar New Year holiday, traditionally a time for family reunions and cultural festivities, has evolved into a new arena for China's technology titans to demonstrate their latest artificial intelligence capabilities. In a concentrated burst of activity during the Spring Festival period, industry leaders Alibaba and ByteDance unveiled significant advancements, highlighting two distinct strategic approaches dominating the global AI landscape: foundational model efficiency and massive, real-world consumer application.

The week surrounding the February 16th New Year's Eve saw both companies launch high-profile initiatives. Alibaba's DAMO Academy made a strategic open-source release of its Qwen3.5-Plus model, claiming a breakthrough in performance-per-parameter that challenges established scaling laws. Simultaneously, ByteDance's Doubao AI assistant integrated deeply with China's most-watched television event, the CCTV Spring Festival Gala, processing a staggering volume of real-time interactions and powering novel visual effects for the broadcast.

This dual display underscores the intensifying competition and rapid maturation of China's AI sector, where technological prowess is being tested not just on academic benchmarks but increasingly in the crucible of large-scale, practical deployment and cost efficiency.

Alibaba's Qwen3.5: A Paradigm Shift Toward "Smarter, Smaller, Cheaper"

On New Year's Eve, Alibaba open-sourced Qwen3.5-Plus, declaring it the new global leader among open-source models with performance "comparable to Gemini 3 Pro." The announcement positioned the release as a capstone achievement for the year. The model's specifications present a compelling case for a shift in industry priorities.

With 397 billion total parameters but only 17 billion activated during inference, Qwen3.5-Plus reportedly outperforms its predecessor, the trillion-parameter Qwen3-Max, while reducing computational deployment costs by 60% and boosting inference speed by eightfold. Alibaba also highlighted its status as a native multimodal model with enhanced Agent capabilities, offered at an aggressive API price of 0.8 RMB per million input tokens.

This move represents a direct challenge to the prevailing "Scaling Law" orthodoxy, which has long equated more compute, data, and parameters with better performance. Industry observers note that this relentless expansion is hitting practical walls: the quadratic computational complexity of the Transformer architecture with longer sequences and the prohibitive cost of inference for enterprise adoption.

"Similar to the Andy-Bill Law in the CPU era, where hardware gains were consumed by more demanding software, the compute dividends from architectural improvements in AI are being depleted by parameter inflation and context window expansion," said a technologist familiar with the development. "Qwen3.5's approach is to extract greater value from every unit of compute."

The technical breakthrough centers on a fundamental innovation within the Transformer architecture itself. The Qwen team integrated its award-winning gated attention mechanism—recognized with a Best Paper award at NeurIPS 2025—into a hybrid architecture combining linear attention and a sparse Mixture-of-Experts (MoE) framework.

In essence, the model employs a learnable "gate" to dynamically evaluate information before the computationally heavy attention calculation occurs. This allows it to decide which segments of data warrant full, precise computation and which can be processed with a lower-cost pathway. This introduces sparsity into the core attention unit of the Transformer, a significant advancement beyond previous MoE models that only sparsified the feed-forward networks.

"Gated Attention addresses a long-standing MoE shortfall," explained an AI researcher. "Previously, attention calculation remained a full, dense operation even in MoE models. This innovation makes attention itself selective, enabling gains in speed and cost without sacrificing the upper bound of model capability."

Native Multimodality and the Agent Ecosystem

The release also marks the Qwen series' first foray into native multimodal design. The model demonstrates superior performance to its larger predecessor across text, image, video, and Agent benchmarks while utilizing less than 40% of the parameters. Industry analysis suggests this stems from a tightly coupled design philosophy where architectural efficiencies enable sophisticated multimodal training, which in turn refines the model's core reasoning.

"Simply stacking modalities and parameters is inefficient. Each new modality exponentially increases compute and memory needs without a proportional gain in reasoning," noted a computer vision specialist. "A native, unified design from the ground up allows for more efficient cross-modal understanding and logical alignment."

Benchmark results cited by Alibaba show Qwen3.5-Plus scoring 87.8 on MMLU-Pro (surpassing GPT-5.2), 88.4 on the GPQA doctoral-level exam (above Claude 4.5), and achieving record scores on instruction-following and Agent-specific evaluations like BFCL-V4. Its visual capabilities also led in several multimodal reasoning and visual question-answering assessments.

These technical advancements are directly linked to Alibaba's broader commercial "playbook." The strategy leverages a virtuous cycle of open-source proliferation, scalable cloud infrastructure, and custom silicon to drive down marginal costs and accelerate adoption.

The company reported that Qwen models have been downloaded over 10 billion times, with over 200,000 derivative models created. This open ecosystem, coupled with deep integration with Alibaba Cloud and its in-house T-Head semiconductor arm, aims to make AI inference radically affordable. The cited price of 0.8 RMB per million tokens for Qwen3.5 is positioned as one-eighteenth the cost of comparable performance from Gemini 3 Pro.

On the application front, Alibaba pointed to the success of its Qwen App, which introduced a consumer AI shopping Agent in January. During the Spring Festival promotions, over 130 million users engaged with the feature, uttering the phrase "Qwen, help me" 5 billion times. The underlying Qwen3.5 model powers enhanced Agent abilities, enabling autonomous operation of mobile and desktop applications for complex, multi-step tasks.

ByteDance's Doubao: Stress-Testing AI at Scale in the Consumer Sphere

While Alibaba focused on foundational model economics, ByteDance demonstrated the power of applied AI at a breathtaking scale through its Doubao assistant's integration with the CCTV Spring Festival Gala. The collaboration served as a live, national-scale stress test for its AI infrastructure.

Doubao's activities during the Gala generated 19 billion AI interactions on New Year's Eve alone. The "Doubao Celebrates the New Year" campaign helped users create over 50 million festive-themed avatar images and generate more than 100 million New Year祝福 messages. The technological peak occurred at 9:46 PM, following a host's prompt for audience interaction, when Doubao's large model handled a peak throughput of 63.3 billion tokens per minute (TPM).

This massive load was managed by ByteDance's cloud service, Volcano Engine, which served as the Gala's exclusive AI cloud partner. The event marked the first time in the Gala's history that AI was used to drive real-time, mass-audience creative interaction.

Beyond interactive features, ByteDance's latest generative video model, Seedance 2.0—which recently gained global attention—was utilized in the production of several Gala performances before its public release. For the program "Song of Riding the Wind," which aimed to animate a treasured ink-wash painting, Seedance 2.0 was tasked with dynamically rendering the artwork's distinctive brushstrokes and ink tones while executing complex motion and camera work to depict running horses naturally.

In another performance, "Celebrating the Flower Goddess," the model showcased detailed control, creating custom, AI-generated visual spectacles for each performer featuring specific flowers and scenes, blending them with physical stage extensions to create a novel narrative structure.

Volcano Engine also deployed other AI technologies for the broadcast. Spatial video techniques created 3D digital doubles of performers that moved in sync with live camera angles and lighting changes. Several embodied intelligent robots appearing on stage were powered by the Doubao model for natural dialogue. Furthermore, Doubao's speech recognition model provided real-time, comprehensive accessibility captions for the entire Gala live stream on Douyin, including hosting segments, cross-talk, and skits that lacked native subtitles.

Divergent Visions, Converging Market Pressure

The simultaneous showcases by Alibaba and ByteDance during the holiday period illuminate the two dominant, and perhaps complementary, thrusts within China's aggressive AI sector. Analysts observe a cultural contrast in development philosophies.

"Silicon Valley firms like OpenAI often position themselves as pioneers defining the future of the technology itself," said a technology strategist based in Shanghai. "In China, with Alibaba's Qwen as a prime example, the drive is intensely pragmatic and industrial—focusing on how to efficiently apply AI to concrete business and consumer scenarios. The Spring Festival activities demonstrate that application demand is now actively pulling model development."

The Qwen strategy, focusing on open-source, cost-leadership, and deep enterprise integration through cloud and chip synergies, aims to become the ubiquitous, economical backbone for AI-powered services. In contrast, ByteDance's Doubao, backed by its immense consumer traffic from platforms like Douyin, is perfecting the art of deploying AI at viral, population-scale for entertainment, creativity, and engagement.

Market data suggests both strategies are gaining traction. According to Omdia, AI-related revenue for Alibaba Cloud has seen triple-digit year-on-year growth for nine consecutive quarters, with Qwen becoming a primary driver for new cloud demand. Sullivan data ranks Qwen first in China's enterprise-level large model API market. Meanwhile, Doubao's successful Gala integration proves the viability of AI for handling billions of real-time requests in a demanding consumer environment.

As the AI "voyage" continues, the competition is increasingly defined not just by who can build the most powerful model, but by who can build the most efficient, scalable, and deployable one. The Lunar New Year showdown indicates that for China's tech giants, the race to dominate the next computing era is being fought simultaneously in the realm of fundamental architectural innovation and in the chaotic, vibrant theater of mass consumer adoption.