During this year's Spring Festival, domestic large models ushered in a wave of intensive updates: Zhipu released a new generation of flagship model GLM-5; Minimax launches Minimax 2.5; DeepSeek has been upgraded on the web and app, and the context length has skyrocketed from 128K to 1 million tokens - equivalent to reading an entire novel at once; Alibaba Tongyi Qianwen's Qwen 3.5 and ByteDance's Doubao Model 2.0 are also expected to be unveiled during the Spring Festival.

These new models generally focus on two major capabilities: programming and agents (i.e., AI agents that can think and perform tasks autonomously).

As the tasks handled by agents become more and more complex, the computing power (tokens) consumed in a single run have risen sharply. If the cost cannot be reduced, the large-scale popularization of AI in the future may be held back by "burning money". Technology runs fast, but whether it can be afforded is the key to landing.

The era of agents is accelerating

On the eve of the Spring Festival, domestic large models ushered in a major upgrade: MiniMax launched M2.5: specially designed for "AI agents".

This is the world's first production-grade model natively designed for agent scenarios, with programming and agent capabilities catching up with the world's top level, benchmarking Anthropic's Claude Opus 4.6. Although the activation parameters are only 10 billion (10B), the efficiency is extremely high: low video memory usage, fast inference speed, and ultra-high throughput of 100 requests per second (100 TPS), far exceeding similar international models.

Zhipu released GLM-5: doubling parameters and leaping performance.

The scale of parameters has expanded from 355 billion to 744 billion in the previous generation, and the activation parameters have also been increased from 32B to 40B, with an average performance increase of more than 20% in programming scenarios such as front-end and back-end development and long task processing. The real programming experience is close to Claude Opus 4.5, and the stock price of Zhipu Hong Kong stocks soared from HK$203 to HK$443 in 4 days, almost doubling, and the market value is close to that of MiniMax.

Agents have become a new "main battlefield" for domestic large model competition

The competition of domestic large models is no longer fighting for who has bigger parameters, but has entered a new stage - comparing technical characteristics, scene landing capabilities, as well as cost and efficiency.

Around the Spring Festival this year, the actions of major manufacturers revolve around this core: Tencent Yuanbao and Alibaba Qianwen attract users through red envelope marketing and strengthen the ecological layout; ByteDance launched Seedance 2.0, DeepSeek upgraded its V4 model, and MiniMax launched the Agent platform

These updates have one thing in common: an increasing focus on "agent" capabilities.

What is an agent?

Simply put, it is a system that allows AI to not only chat and answer questions, but also actively think, call tools, and perform tasks. For example, automatic ticket booking, writing code, doing research, generating PPT, etc.

General large models are often "powerless" in professional and complex business scenarios. By integrating professional knowledge, calling external tools, and orchestrating workflows, agents can truly penetrate vertical fields such as finance, healthcare, and manufacturing to provide automated and high-value solutions.

This also reflects the change in market mentality: people no longer pay for "how good AI can talk about", but are more concerned - can it help me work, improve efficiency, and save money in real work?

| The large-scale implementation of AI still has to pass the cost hurdle

Despite the rapid development of domestic AI, there are still several practical challenges to truly implement on a large scale:

1. Ecological "fragmentation" and lack of unified standards

From chips and development frameworks to large models and applications, all links in the domestic AI industry chain are "fighting separately", with incompatible interfaces and non-interoperable tools, which increases the complexity of development and deployment, and a more unified ecosystem is urgently needed.

2. The cost is too high, and the enterprise cannot calculate the "return account"

Although it is becoming cheaper to call APIs, many companies still want to privatize AI models locally for security or compliance reasons; This means investing a lot of money to buy servers, build computer rooms, and maintain an operation and maintenance team for a long time; If it is not clearly proven that AI can bring quantifiable business benefits (such as cost reduction and revenue increase), it is difficult for enterprises to make up their minds to invest.

What's even more troublesome is that as agent tasks become more and more complex - such as automatically completing a market report, you need to check data, write analysis, and make charts

The computing power (token) consumed by a single task has skyrocketed, and the cost remains high.

If it cannot be reduced, AI agents can only be used in "high-value, low-frequency" scenarios, and it is difficult to popularize them in daily business.

3. Insufficient reliability, "hallucination" is still a hard wound

In industries with extremely high accuracy requirements, such as finance, medical care, and manufacturing, once AI "talks nonsense" (i.e., "hallucinates"), the consequences may be serious;

When the current model performs multi-step complex tasks, errors are easy to accumulate step by step, which eventually leads to the failure of the entire process. Without sufficient reliability, enterprises do not dare to let AI truly "work".

4. Governance and trust mechanisms have not kept up

When AI moves from "auxiliary suggestions" to "autonomous execution", new problems arise:

How to manage permissions? What systems can AI operate? Who is responsible for what went wrong? Is it the developer, the enterprise, or the AI itself? Can the operation process be audited and traced?

Only by solving these governance problems can enterprises really dare to hand over key tasks to AI.

5. Data security and privacy remain sensitive thresholds

Especially in the fields of government affairs, finance, and medical care: enterprises are worried that data uploaded to the public cloud will be leaked "out of the domain"; Whether the model training data is compliant and biased is also worrying; How to protect privacy in the process of agent interaction with users is still a dual problem of technology + system.

Embrace AI that can be "landed" and be wary of the "pure parameter" bubble

For A-shares/Hong Kong technology stocks:

Priority is given to companies with vertical scenario agent capabilities: such as AI solution providers in finance, medical care, government affairs, manufacturing and other fields; Pay attention to the collaborative ecology of domestic computing power + models: Huawei Ascend, Cambrian, iFLYTEK, Zhipu AI (Hong Kong stocks), etc.; Be cautious about pure large model concept stocks - if there are no real customers, no cost control, and no governance framework, it is difficult to escape the question of "PPT AI".

For primary markets/entrepreneurs:

Cutting into high-value, low-fault tolerance scenarios requires caution (such as medical diagnosis, industrial control); It is more suitable for starting from low- and medium-risk areas such as office efficiency improvement, customer service automation, and marketing generation; The core capability is not "how big is the model", but "task closed-loop + cost controllable + explainable".

For Enterprise Users:

SaaS-based agent services (such as MiniMax and Alibaba Cloud Bailian platform) can be piloted to avoid heavy asset investment; Key evaluations: error rate, audit capability, and data not leaving the domain