China’s major AI models kick off the “Spring Festival season”! Awaiting breakout consumer AI hits

2026’s Spring Festival is no longer just a carnival of consumer spending, but has transformed into a race of speed and passion among China’s AI giants vying for supremacy on mobile platforms. According to Chase Wind Trading Desk, JPMorgan Chase noted in its latest February 11 research report that China’s internet and AI industries are experiencing the densest wave of flagship model launches in history. This is no longer a solo show for a single model, but a game of musical chairs to see who can convert “technological spillover” into “blockbuster consumer products” the fastest. Crowded “Spring Festival Release Season” DeepSeek’s surprise Spring Festival strategy in 2025 has become an industry textbook. By 2026, everyone has learned from it. According to JPMorgan’s observation, the market is once again prepared for a similar “holiday season release,” but differs in one key way: “This time it isn’t a lone standout. We see a large number of flagship and near-flagship updates congregating in the same window.” - ByteDance was first to play its hand, unveiling a three-model “gift package”: Seedance 2.0 (video), Seedream 5.0 (image), and Doubao 2.0. Among these, Seedance 2.0 has shown “blockbuster” potential. - Alibaba wouldn’t be left behind. Reports say it’s set to launch Qwen 3.5 in mid-February, backed by a 3 billion yuan incentive program to attract users. - Zhipu launched GLM-5 (formerly the mysterious “Pony Alpha”) on February 11, expanding parameter size from 355 billion to 744 billion. - DeepSeek’s V4 version is reportedly targeting mid-February, with key improvements in coding and ultra-long prompt processing. On February 11, reports indicated DeepSeek launched a new model supporting up to 1M tokens in context. - MiniMax’s latest large model M2.5 went live on the Agent platform on February 11. JPMorgan bluntly states that such cluster launches will result in “winner-takes-all” dynamics: “When several models appear simultaneously, developers and companies ramp up comparative testing… those that lag behind stand to lose the most.” During the attention-scarce Spring Festival window, if a lab can’t produce a credible flagship update, it may instantly drop off developers’ trial lists. “In the Spring Festival window, attention is scarce, but in a way, plentiful: users try more things, yet decide more quickly which products they’ll keep using. That’s why we see Spring Festival as a leaderboard reset period.” DeepSeek V4 and the Ambition for “Cost Reduction” The market’s most watched variable remains DeepSeek. JPMorgan analysts point out that DeepSeek’s potential Spring Festival release isn’t mainly about the chatbot itself, but the “platform economic effect” it may unleash. DeepSeek’s latest paper, “Conditional Memory Based on Scalable Lookup,” reveals its technical path: improving quality via “conditional memory” as a second sparse axis, without brute-force computational upgrades. The report suggests that DeepSeek’s proposal to add “conditional memory” beyond Mixture-of-Experts (MoE) might be key to next-generation models. It’s not just about lighter models, but shifting expensive dense calculations to cheaper retrieval operations. “If the technology performs as described, it will mean improved efficiency—not just model lightweighting, but moving part of what the model ‘does’ from costly dense computation to more affordable retrieval/memory operations.” “If it achieves efficiency-first and cost-reduced inference, AI can be economically embedded directly in high-frequency consumer products, not just as standalone chatbots.” This means AI will shift from expensive “toys” to cheap “tools.” If inference unit costs drop, spending more tokens on multi-step reasoning and tool execution becomes reasonable, which will push China’s AI adoption curve from “chat tools” towards “AI-embedded” as the default interface in high-frequency apps. Who Wins? Tencent May Be the Biggest Beneficiary In this model war, JPMorgan offers a counterintuitive take: the biggest winners may not be the model makers, but Tencent. The logic is clear: Tencent owns WeChat and QQ, China’s highest-frequency communication interfaces. “We believe Tencent stands to be the biggest beneficiary from DeepSeek’s potential new releases… Tencent has already begun integrating third-party model capabilities into WeChat and other core consumer platforms, increasing the likelihood that stronger model performance will enhance user experience. If model quality improves significantly, we expect Tencent to confidently expand and deepen AI integrations across its high-frequency communication interfaces, including QQ.” For Alibaba and Baidu, the situation is more complex. Stronger models can improve their product user experience (e.g., Taobao Q&A, Baidu Search); but if DeepSeek sparks another “price war,” the entire industry’s API service will face intense deflationary pressure. “If newly released DeepSeek models substantially improve efficiency, we expect the industry-wide model API service will continue to face price pressure, possibly impacting short-term unit economics.” As for vertical players like Trip.com, Beike, and Kuaishou, JPMorgan believes this is unequivocally positive. Powerful open-source models lower their technical barrier, allowing them to use proprietary data for fine-tuning at lower cost, accelerating product iteration. Waiting for a Consumer AI Blockbuster Although there’s strong market enthusiasm, JPMorgan remains objectively sober. At present, “consumer-facing agents” often look impressive in demos but struggle with inconsistency when implemented. The Spring Festival’s mass user test will be the true litmus test. “We agree with current market skepticism: Many 2C agents’ demos remain inconsistent, and consumer satisfaction typically falls short when mass users try them on real tasks.” But if this season’s flagship models truly solve reliability and latency in tool usage, the inflection point may finally arrive. As JPMorgan puts it: “True adoption signals aren’t the hype on release day, but whether incumbents let AI become built-in default functions on high-frequency interfaces—because that’s what drives sustained reasoning demand.” Valuation Rebuild: Looking Toward 2030 Profitability Additionally, for secondary market strategy, JPMorgan maintains “overweight” ratings for pure model developers Zhipu and MiniMax. Zhipu’s GLM-5 is open-source SOTA among agents, ranking first on multiple benchmarks. MiniMax, through a full-spectrum of models (text, video, audio), has achieved dual B2B/B2C commercialization. JPMorgan’s valuation logic now skips short-term losses, looking straight at 2030: - Zhipu: Target price HK$400, based on 30x 2030 forecast P/E ratio. - MiniMax: Target price HK$700, based on 30x 2030 forecast P/E ratio. Analysts believe that as model capabilities approach the global frontier, the valuation logic will fundamentally change: “Valuation upgrades won’t be so much about national pride but will focus on economic utility: stronger willingness to pay, and higher API workload retention rates.” ~~~~~~~~~~~~~~~~~~~~~~~~ The above content is courtesy of [Chase Wind Trading Desk](https://mp.weixin.qq.com/s/uua05g5qk-N2J7h91pyqxQ). For more detailed analysis, including real-time interpretation and frontline research, please join [Chase Wind Trading Desk Annual Membership](https://wallstreetcn.com/shop/item/1000309). Risk Warning and Disclaimer The market carries risks; investments must be made cautiously. This article does not constitute personal investment advice, nor does it consider the unique investment goals, financial situations, or needs of any individual user. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Invest at your own risk.