National Model Crisis and Opportunity

In the past few years, Guomo has made rapid progress in engineering efficiency and cost control. But if it can only provide near-cutting-edge capabilities at a lower price, can this business continue to support high-cost basic model research and development? This is where the market worries come from.

At the same time, the large model industry is undergoing an important round of changes. From Coding to Agent, and then to Computer Use, models no longer just answer questions, but begin to enter real environments and complete complex tasks. Training methods, data requirements and business models are all changing accordingly.

For Chinese model companies, this means new challenges and may also be opportunities to re-establish competitive advantages.

1. There are still new solutions to the upper bound of intelligence

In the past few years, innovations in engineering efficiency of Chinese models have profoundly changed the cost structure of global large models. MoE, sparse attention, low-precision training, inference acceleration, and software and hardware collaboration enable increasingly powerful models to be trained and deployed at lower costs. But there is another more important issue in model development: how to continue to improve the upper bound of intelligence under limited computing resources?

Chinese companies are exploring different answers.

DeepSeek pays more attention to changing the computational structure of the model. DeepSeek V4.1 Flash released in September adopts a 552B parameter MoE architecture and introduces an asymmetric Causal Encoder–Decoder design. Only about 8B parameters are activated on the input side and about 16B on the output side. Compared with the previous generation, KV Cache’s demand for high-bandwidth video memory has been reduced to one-quarter, and the demand for SSD storage has been reduced to one-eighth.

This kind of architectural innovation not only reduces the cost of inference, but also opens up new efficiency space for long context processing and the training and deployment of larger models.

The technological accumulation of wisdom spectrum can be traced back even further. In 2022, the Zhipu and Tsinghua teams open sourced GLM-130B, a Chinese-English bilingual pre-training model with 130 billion parameters. At that time, there were not many research teams that could complete model training with hundreds of billions of parameters and open weights.

The GLM-130B paper records issues such as Loss Spike and training divergence in large-scale training, as well as the technical solutions used to stabilize training. The team also studied INT4 quantization so that the 100 billion parameter model can run on 4 RTX 3090s.

In August this year, Zhipu released GLM-5.3-Flash, redesigning the base architecture and training plan.

The total parameters of the model are 320B, and each Token only activates 18B parameters. It adopts a 45-layer network structure, introduces a hybrid architecture of sparse attention and linear attention, and Manifold-Constrained Hyper-Connections, and uses 30 trillion Token multi-modal pre-training data.

According to official disclosures, compared with GLM-5.3, the new architecture reduces the amount of attention calculation by about two-thirds and the KV Cache size by about 77%, while maintaining strong coding and agent capabilities.

From stable training with hundreds of billions of parameters to hybrid attention and computing structure optimization, Zhipu continues to study how to achieve higher effective computing efficiency for the model.

Alibaba Qianwen continues to invest on a larger scale. The total parameters of Qwen3.8-Max reach 2.4 trillion, combining million-level context, multi-modal capabilities and Alibaba Cloud infrastructure to explore the capability boundaries of larger models.

These routes are not in conflict. Larger model scale, more efficient computing architecture, and better training data may lead to the next round of capability breakthroughs.

Especially as trillion-parameter MoE becomes an important technical route, there are still a lot of issues worthy of study in stable pre-training itself: Loss Spike, expert load imbalance, route collapse, attention logit anomaly, hyperparameter migration of new optimizers, and the design of data ratio, mid-training and learning rate annealing. These details affect the model’s final knowledge representation, reasoning potential, and scalability.

Since this year, Anthropic and OpenAI have continued to strengthen model access control and anti-distillation measures, further increasing the importance of independent research. Distillation can help student models efficiently transfer existing capabilities, but to explore new tasks for which there are no mature answers, you still ultimately need your own data, training methods, and experimental systems.

Guomo has accumulated solid engineering capabilities. In the next stage, what is more worth looking forward to is the combination of these capabilities and basic research. One of the most important resources is talent. People who can ask the right questions, design key experiments, and judge whether the increase in ability is a real signal or evaluation noise are even rarer than those who are simply familiar with the training process.

China has a huge pool of algorithmic and engineering talents. As the research and development of basic models continues to deepen, a group of teams that have long-term research on model architecture, optimization algorithms and large-scale training are becoming the basis for the next round of technological innovation. In addition to pre-training, reinforcement learning is opening up another ability growth curve.

2. Post-training is opening up new technical space

In the past, large models were mainly learned from static corpora. Entering the Agent era, models begin to change the environment through actions and then learn from environmental feedback.

A Coding Agent not only needs to generate code, but also needs to read the warehouse, install dependencies, run tests, modify files, and continuously adjust according to the execution results.

This training method puts forward new requirements for the environment, rewards and algorithms. The environment must be runnable, reproducible, reconfigurable, and capable of supporting large-scale parallel training. The difficulty of the task should also be close to the current capability boundary of the model: if it is too simple, there will be no learning gradient, and if it is too difficult, it will be difficult to obtain effective feedback.

獎勵Task completion needs to be defined as precisely as possible. Programming and mathematics can use unit testing, compilation results and answer verification; open tasks such as finance, law, and research require rubrics, generative reward models, process rewards, and multi-layer verification. They must also prevent model modification tests, bypass verification, and other reward hacking behaviors.

算法It is necessary to solve the credit allocation in long-range tasks, the Off-policy bias of asynchronous RL, entropy collapse, insufficient exploration, and the context compression and memory problems in long trajectories. There is still no generally accepted complete answer to these difficult questions, which also means new research space.

Kimi and Zhipu are making progress in this direction.

Kimi K3.1 Pro, released in October, further expands reinforcement learning calculations based on Kimi K3, and optimizes the RL framework, Sandbox, Rollout scheduling and environmental quality inspection. In the Terminal-Bench 4.0 evaluation of Artificial Analysis, its score reached 67%, which was close to the strongest closed-source model at the time.

GLM-5.3 follows the GLM-5.2 base and invests its main capabilities in long-range mission environments and reinforcement learning.

According to the technical report, its training system extracts task patterns from real work through the research agent, automatically constructs an executable environment, and then uses the judge agent to check whether the task can be completed, and discovers reward vulnerabilities through validators and multi-layer inspections.

The training task has been extended to machine learning infrastructure optimization: the model needs to analyze computing clusters, storage systems, codes and experimental results, find performance bottlenecks, complete optimization and verify actual effects.

Zhipu disclosed that the end-to-end completion performance of GLM-5.3 on the internal Code Bench is about 50% higher than that of GLM-5.2; through asynchronous scheduling and Rollout optimization, the long-range Coding RL training throughput is increased by more than 2.3 times.

These developments show that the growth in capabilities of large models is forming a new technology system: base models provide basic capabilities, complex environments generate training tasks, reward systems provide feedback, and reinforcement learning continues to improve task completion capabilities.

Even more promising is the possibility of the model participating in its own training. From human construction tasks and model learning tasks, it gradually moves towards the model itself generating tasks, building environments, executing experiments, verifying results, and then using feedback to improve its capabilities. This route still faces issues such as validator reliability, self-reinforcing bias, and reward vulnerabilities. Especially when the generator’s power exceeds that of the validator, the model may continuously reinforce errors. However, if a reliable closed loop of generation, verification, and learning can be established, important changes may occur in the efficiency and capability growth of large model R&D.

3. The next upgrade of the data industry may be in China.

The development of reinforcement learning may also redefine the data industry.

China has a mature data service supply chain, a large number of software engineers and rich industrial scenarios. But the new generation of Agent training requires more than just text annotation and manual evaluation.

The deliverables of high-quality training data are turning toExecutable environments, validators, and expert tracks. For example, training a financial agent requires constructing a workspace containing multiple financial reports, data conflicts and business constraints, allowing the model to complete verification, calculation, analysis and delivery, and then verify the task results.

This requires data service providers to not only understand the business, but also have software engineering, Sandbox construction, Reward design, and even basic reinforcement learning experimental capabilities to determine whether the environment can continue to generate effective learning signals.

Companies such as Mercor that have built their businesses around professional knowledge and high-quality AI training data have emerged overseas. China also has the opportunity to cultivate a number of technology-intensive data service companies.

There are at least two paths: traditional data companies introduce RL and Agent talents and upgrade to training environment suppliers; existing RL and Agent companies use software engineering and industry experience to provide complex task environments to basic model companies. This may lead to the formation of a new business model: Environment as a Service.

What customers buy is no longer just the number of data items, but the environment, validators and effective learning signals that can continuously train the model. Furthermore, the environment construction itself is being automated.

The technical report of GLM-5.3 has demonstrated the process of automatically generating the task environment and part of the verification mechanism. As this capability expands, the cost of producing a standardized environment is likely to decline rapidly.

Human value will be further concentrated in areas that cannot be covered by automatic generators: the long tail of real business, private enterprise scenarios, complex industry knowledge, and the ability to evaluate the quality of automatically generated environments.

In the future, there will even be an “environment of environment” that specifically evaluates the training environment to determine whether the difficulty of the task is reasonable, whether there are loopholes in the rewards, and what the model can learn from it.

China’s manufacturing, finance, software development, e-commerce and professional services sectors have a large number of complex and specific work processes. In the past, these scenarios were viewed more as AI application markets. In the Agent era, they may also become training resources for the next generation of models.

Whether real industry knowledge can be transformed into a reproducible, verifiable, and scalable learning environment will determine how much value this resource advantage can release.

4. Large models are starting to work, and opportunities in the Chinese market are expanding.

Training methods change because the model’s job is changing.

From Coding to Agent, and then to Computer Use, models gradually enter the computer environment and assume more and more complex tasks. Coding was the first to achieve breakthroughs, because the code can be executed and tested, and can form relatively clear training feedback.

Agents require multi-step decision-making, heterogeneous tool invocation, external state maintenance and long-term planning, and rewards are often sparse and delayed.

Computer Use further requires the model to understand the GUI, complete visual grounding, and take actions in the ever-changing software interface.

Every step forward requires enough basic data, a broad enough training environment, complex enough tasks, and accurate enough trajectories and feedback.

Beyond computers, there are opportunities to interact with the physical world.

Robots, industrial controls, automation equipment, and laboratory systems all require models to understand their environment, perform actions, and adjust based on feedback. The physical environment is difficult to replicate, and training costs and safety requirements are higher, but it also corresponds to a broader industrial space. These changes are changing the business value of the model.

In the past, users mainly purchased intelligent services based on tokens; now, more and more companies are beginning to measure AI based on final task results.

A pricier model may also be more economical than a lower-priced model if it can get the job done in fewer steps.

The credit consumption problem that occurred after Kimi K3.1 Pro went online reflects this change: the capabilities of complex agents have improved, but sub-agent calling and long-range reasoning have also increased usage costs. How to deliver high-quality tasks at a reasonable cost will become an important competitive indicator in the next stage.

Different Chinese companies are looking for their own market position.

MiniMax starts with voice, video and AI native applications, and gradually strengthens Coding, Agent and Computer Use. The M3.1 Flash released in October focuses on optimizing inference speed and cost. Officially disclosed that the cost has dropped by about 70% compared with the previous version.

In the first half of 2026, MiniMax’s revenue reached US$116.6 million, a year-on-year increase of 283.1%; of which revenue from open platforms and enterprise services was US$73.9 million, accounting for 63.4% of total revenue.

Qianwen and Hunyuan have the ecological advantages of large technology companies. Alibaba can enter enterprise and office processes through Alibaba Cloud and QwenWork; Tencent integrates hybrid capabilities into software development and productivity scenarios through Tencent Cloud, CodeBuddy, and WorkBuddy.

The Hy4 preview released by Tencent in August has a total parameter of 770B and an activation parameter of 49B, focusing on programming, office and scientific research. Tencent also disclosed that the model improved end-to-end throughput by 31.8% in internal experiments by analyzing and optimizing its own inference infrastructure.

These paths illustrate that the value of AI is extending from model APIs to software, tools and complete business processes.

China has both independent model companies, a huge software and Internet ecosystem, and complex scenarios in manufacturing and professional services. As Agents truly begin to take on work, more new combinations will emerge between model capabilities, system engineering, and industry knowledge.

5. From global usage to global value

If technology and applications create new opportunities, then commercialization needs to answer how to make these opportunities support long-term research and development.

Anthropic’s prospectus has proven that high-value enterprise customers are willing to continue paying for reliable intelligent services. But this growth also requires huge investment: in the first three quarters of 2026, its computing power and infrastructure spending reached 18.808 billion U.S. dollars, and it signed a large-scale long-term computing power purchase agreement.

One of the current advantages of the Chinese model is that it has entered the global developer ecosystem through open weighting, efficiency innovation and price competition.

The next opportunity is to establish a more diversified business model: basic model API, subscription products, enterprise services, reasoning infrastructure, and agent products for specific tasks.

Zhipu’s recent OpenRouter performance provides a case worth observing.

According to a set of statistical snapshots in early November, GLM-5.3 Flash has called approximately 15 trillion Tokens in the past seven days, ranking second on the site, with a month-on-month increase of 13%; approximately 55.1 trillion Tokens have been called in the past 30 days, which is basically the same as the previous 30 days.

GLM-5.3 ranked tenth in both the past 7 days and the past 30 days, with a call volume of approximately 19.1 trillion Tokens in the past 30 days, a month-on-month increase of 54%. In the second month after the model was released, usage was higher than in the first month.

Even more interesting is the price. The price of GLM-5.3’s OpenRouter is about 7 to 9 times that of DeepSeek V4.1 Flash 1017, and the token usage is about one-eleventh of the latter.

This cannot directly deduce the gap in model capabilities, but it shows that Chinese models are forming different price and usage combinations. For some developers, the model’s complex task capabilities, long-term reliability, and overall delivery cost may be more important than the price of a single token.

Zhipu’s financial data also shows changes in its business model.

In the first half of 2026, Zhipu’s revenue reached 954 million yuan, a year-on-year increase of 399.7%; of which open platform and API revenue was 825 million yuan, a year-on-year increase of approximately 2736%, accounting for 86.5% of total revenue.

Compared with the business structure that was dominated by localized deployment in the past, standardized model calling has become the main source of revenue.

Another advancement comes from inference infrastructure. Zhipu disclosed that it has achieved model inference on a 100,000-level domestic chip scale, and through system optimization has reduced the unit Token inference cost by about 80% compared with the beginning of the year. From base architecture and post-training to inference costs and API revenue, technical efficiency is directly affecting the commercial capabilities of model companies. But opening up weights also brings new business issues.

After the model is released, third-party inference service providers can deploy the same model to compete with the original R&D enterprise. Epoch AI has observed that about 20 days after GLM-5.3 Flash opened its weight, the share of OpenRouter Tokens undertaken by Zhipu itself for this model dropped from 88% to 22%.

Growth in model usage does not mean that all commercial revenue will flow to the original R&D company.

Therefore, the competition for open source models in the future will not only expand developer coverage, but also include whether they can form a stable source of income through better model capabilities, product experience, inference services, and corporate relationships.

DeepSeek expands the developer ecosystem with architectural innovation and high cost performance; Kimi invests in higher intelligence limits and complex tasks; Zhipu tries to connect basic research, post-training, domestic reasoning and API commercialization; MiniMax also operates global AI products and enterprise services; Alibaba and Tencent can obtain model value through existing cloud computing and software businesses.

There has been real progress along different paths.

As AI shifts from selling models to providing work results, Chinese companies also have the opportunity to build business capabilities that are not entirely dependent on low-price competition.

6. In the next round of competition, China needs more teams like this

Opportunities are emerging, but challenges certainly remain.

The research and development of cutting-edge models still relies heavily on computing power, and the progress of domestic chips on the inference side needs to be further extended to large-scale training and a complete software ecosystem.

A high-quality training environment cannot be automatically obtained with industry data. Environmental construction, reward verification and professional task design require closer industrial cooperation.

At the same time, as the Agent enters the real working system, permission control, data security, reward cheating prevention, process auditing and manual takeover must all become part of the model capabilities.

There is also R&D investment. In the first half of 2026, Zhipu’s R&D expenses reached 2.131 billion yuan, with a loss of approximately 2.072 billion yuan during the period; MiniMax’s adjusted net loss during the same period reached US$293 million. Commercialization has begun to grow, and basic model research and development still requires continued large-scale investment.

A new architecture may require multiple rounds of experiments to be verified, a batch of RL environments may fail due to reward design issues, and a large-scale training may also suffer losses due to system instability.

Basic research requires such a trial and error process.

The situation of Chinese model companies in the past few years is somewhat like “Xiaomi plus rifle”: the computing power and capital are not as sufficient as their competitors, but the research direction must continue to follow the world’s cutting-edge. Limited resources need to support technology research and development, engineering optimization, product delivery and commercialization exploration at the same time.

But it is precisely under such conditions that China has formed a group of teams that can independently carry out large-scale base training, research model architecture, construct reinforcement learning environments, build reasoning infrastructure, and provide models to developers around the world.

From 100 billion parameter pre-training in 2022 to today’s trillion parameter architecture, long-range Agent and automatic training environment, technology accumulation is no longer limited to a certain model version.

For the next stage of China’s AI industry, these research teams, engineering experience and developer ecosystem may be more important assets than a single ranking on the list.

China also needs more patient capital that can ride through cycles to provide stable support for long-term basic research, allowing researchers to have space to explore directions that have not yet been verified.

Because the competition for large models is far from over.

How the upper bound of intelligence can continue to improve, whether reinforcement learning can develop a more effective training paradigm, how models can autonomously generate and verify learning tasks, how agents can enter real production processes, and how intelligence can further interact with the physical world are all still open questions. In these directions, China already has its own research accumulation, engineering capabilities, talent network and industrial scenarios.

What China’s large model industry needs is not just one or two companies that are leading on a certain list, but a group of teams that can continue to explore in different technical directions.

DeepSeek, Kimi, Zhipu, MiniMax, Qianwen, Hunyuan…the routes they each choose are also expanding the way China participates in global basic model competition. They are the vanguard of China’s large model industry. These companies don’t necessarily get it right every time, nor do they necessarily lead to the same end point. But they are already using their own research and engineering practices to answer questions for which there are no standard answers in the next generation of smart technologies.

In the past few years, the Chinese model has proven that advanced intelligence can be achieved with higher efficiency and can also serve global developers through an open ecosystem. In the future, China will have the opportunity to form its own technological advantages in new model architectures, training paradigms, agents and real industrial applications.

From catching up with existing intelligence to exploring new ways of growing intelligence, China’s big models are moving towards a broader technological space. The next chapter of this competition is still worth looking forward to.

———————————————

Related Reading

• Nvidia live updates: Jensen Huang ousted, CFO Colette Kress to serve as interim CEO2026-11-13
• Are mangoes not afraid of red fruits?2026-11-12
• Original: After Huang Renxun, who is most likely to take over Nvidia?2026-11-12
• Samsung’s roll-screen mobile phone is exposed, creating a new form ahead of Huawei and Apple, and folding screens are expected to be eliminated2026-11-12
• Musk’s favorite car is going to be discontinued.2026-11-12