share_log

China’s large-scale model 'Spring Festival release' kicks off! Awaiting a blockbuster in consumer-grade AI.

wallstreetcn ·  Feb 12 10:00

DeepSeek's Spring Festival surprise tactics in 2025 have become an industry textbook. By 2026, everyone had learned this move.

ByteDance, Alibaba, Zhipu, and DeepSeek launched flagship models during the 'Spring Festival season,' with competition shifting from pure model performance to the implementation of 'efficiency' and 'intelligent agents.' Analysts suggest that if DeepSeek V4 achieves cost reduction in reasoning, it will drive AI integration from standalone apps into high-frequency applications such as WeChat, making Tencent a major beneficiary.

The Spring Festival of 2026 is no longer just a consumer frenzy but has evolved into a race among China's AI giants to seize the 'mobile entry point' with speed and passion.

According to TradingView, in its latest research report released on February 11, JPMorgan pointed out that China’s internet and AI industries are witnessing the most intensive wave of flagship model releases in history. This is no longer a one-model show but a game of musical chairs about who can fastest turn 'technological spillover' into a 'consumer-level blockbuster'.

The Crowded 'Spring Festival Release Season'

DeepSeek’s surprise tactics during the Spring Festival of 2025 have become an industry textbook. In 2026, everyone has learned this move.

According to JPMorgan’s observations, the market is once again prepared for a similar 'holiday release season,' but with one key difference: 'This time it’s not a solo performance; we see numerous flagship and near-flagship updates converging in the same window.'

ByteDance made the first move, unveiling a three-model 'package': Seedance 2.0 (video), Seedream 5.0 (image), and DouBao 2.0. Among these, Seedance 2.0 has already shown signs of becoming a 'blockbuster.'

$BABA-W (09988.HK)$ Not to be outdone, reports indicate preparations to launch Qwen 3.5 in mid-February, accompanied by a 3 billion yuan incentive plan to boost customer acquisition.

$KNOWLEDGE ATLAS (02513.HK)$ Released GLM-5 (previously known as the mysterious 'Pony Alpha') on February 11, expanding the parameter scale from 355 billion to 744 billion.

DeepSeek was reported to aim for a mid-February release of its V4 version, focusing on improvements in coding and ultra-long prompt handling. On February 11, reports emerged that DeepSeek had updated its new model to support a maximum context length of 1 million tokens.

$MINIMAX-WP (00100.HK)$ The latest large model M2.5 was launched on the Agent platform on February 11.

JPMorgan stated outright that this clustered release will lead to a 'winner-takes-all' scenario:

"When multiple models appear simultaneously, developers and enterprises will intensify comparative testing... the adverse impact on underperformers will be more pronounced."

During the attention-scarce Spring Festival window, if a lab fails to deliver a credible flagship update, it might drop off the list of developers' trial considerations altogether.

"During the Spring Festival window, attention is scarce but in a sense abundant: users will try more things, but they will also decide more quickly which products to continue using. This is why we view the Spring Festival phase as a reset period for rankings."

DeepSeek V4 and the Ambition of 'Cost Reduction'

The variable the market is most focused on remains DeepSeek.

JPMorgan analysts noted that the core impact of DeepSeek's potential Spring Festival release lies not in the chatbot itself, but in the possible unleashing of 'platform economic benefits'.

DeepSeek's latest paper, 'Conditional Memory Based on Scalable Lookup', reveals its technical approach: improving quality by using 'conditional memory' as a second sparse axis without resorting to brute-force computational upgrades.

The research report argues that DeepSeek’s proposed technique of adding 'conditional memory' outside of Mixture-of-Experts (MoE) models could become key for the next generation of models. This is not just about making the model lighter; it shifts expensive dense computation to cheaper retrieval operations.

"If the technology implementation indeed aligns with what the paper describes, this would imply efficiency improvements—not merely a simplification of the model; it essentially transfers part of what the model 'does' from costly dense computation to cheaper retrieval/memory operations."

If it prioritizes efficiency and achieves cost reduction in reasoning, artificial intelligence can be economically embedded directly into high-frequency consumer products, rather than remaining standalone chatbots.

This means that AI will evolve from an expensive 'toy' to an affordable 'tool.' If the unit cost of reasoning decreases, spending more tokens on multi-step reasoning and tool execution becomes reasonable, which will push the domestic AI adoption curve from 'dialogue tools' toward the highest-frequency product interface: 'AI integration.'

Who are the winners? The biggest beneficiary may be Tencent.

In this model competition, JPMorgan offered a counterintuitive analysis: the biggest beneficiary may not be the model vendors but rather others.$TENCENT (00700.HK)$

The logic is very clear: Tencent owns WeChat and QQ, the two highest-frequency communication interfaces in China.

We believe Tencent is well-positioned to become the biggest beneficiary of DeepSeek's potential new product releases...

$TENCENT (00700.HK)$Has started integrating third-party model capabilities into core consumer entry points such as WeChat and Yuanbao, increasing the likelihood that stronger model performance will translate into tangible user experience improvements.

If model quality improves significantly, we also expect Tencent to expand and deepen AI functionality integration across high-frequency communication interfaces, including QQ, with greater confidence.

For$BABA-W (09988.HK)$and$BIDU-SW (09888.HK)$The situation is more complex for some. On one hand, stronger models can enhance the user experience of their products (e.g., Taobao Q&A, Baidu Search); on the other hand, if DeepSeek triggers another 'price war,' the entire industry’s API services could face significant deflationary pressure.

If the newly released DeepSeek significantly improves efficiency, we expect the entire industry’s model application programming interface services to continue facing price pressures, which could strain short-term unit economics.

As for$TRIP.COM-S (09961.HK)$$BEKE-W (02423.HK)$$KUAISHOU-W (01024.HK)$Vertical giants like Trip.Com, according to JPMorgan, this is absolutely positive news. Powerful open-source models lower their technical barriers, enabling them to fine-tune proprietary data at a lower cost and accelerate product iteration.

Awaiting the 'blockbuster' validation of consumer-grade AI

Despite the capital market's enthusiasm, JPMorgan has maintained an objective calmness.

Current '2C agent' demonstrations often appear impressive but frequently encounter inconsistency issues when implemented. The large-scale user testing during the Spring Festival will serve as a true litmus test.

We share the market’s current skepticism: many ‘2C agent’ demonstrations remain inconsistent, and consumer satisfaction typically falls below expectations once users apply them to real-world tasks on a large scale.

However, if the flagship model released during this year’s Spring Festival truly resolves issues related to the reliability and latency of tool usage, the turning point may indeed have arrived.

As JPMorgan concluded: 'The true signal of adoption is not the noise on launch day, but whether existing giants will make AI a default high-frequency interface feature, as this is what drives sustained demand for reasoning.'

Valuation Restructuring: Focusing on Profitability by 2030

In addition, in terms of secondary market strategies, JPMorgan has approached pure model developers$KNOWLEDGE ATLAS (02513.HK)$and$MINIMAX-WP (00100.HK)$Maintain a “Buy” rating.

Zhipu's GLM-5 achieved open-source SOTA (State-of-the-Art) in Agent capabilities and ranked first in multiple evaluation benchmarks. MiniMax, with its full-spectrum models (text, video, audio), has realized dual commercialization in both B2B and B2C sectors.

JPMorgan's valuation logic has bypassed short-term losses and is directly focused on 2030:

  • Zhipu: Target price of 400 HKD, based on a 30-times forward price-to-earnings ratio for 2030.

  • MiniMax: Target price of 700 HKD, based on a 30-times forward price-to-earnings ratio for 2030.

Analysts believe that as model capabilities approach the global frontier, the rationale for valuation upgrades will undergo a qualitative change:

“The increase in valuation may no longer be so much about national pride but rather more focused on economic benefits: stronger willingness to pay and higher API workload retention rates.”

Editor/Melody

The translation is provided by third-party software.


The above content is for informational or educational purposes only and does not constitute any investment advice related to Futu. Although we strive to ensure the truthfulness, accuracy, and originality of all such content, we cannot guarantee it.
    Write a comment