← 返回

2026.09.24 | TB | LLM | China vs Overseas Models & Agents

原标题：AI Foundation Model Industry 2026 Series – Chinese & Overseas Models & Agent Product Architecture Updates (Conducted in Mandarin)

#llm #agents

Third Bridge · 2026-09-24

Analyst-led • 17 Sep 2026

Companies: 2513 HK, 700 HK, META US, BABA US, 100 HK

Expert CEO at Shenzhen Ruoxi Technology Co Ltd

Eric Shen (ES) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

ES: Hello and welcome to Third Bridge’s Interview entitled AI Foundation Model Industry 2026 Series – Chinese & Overseas Models & Agent Product Architecture Updates (Conducted in Mandarin). My name is Eric Shen, Third Bridge’s Sector Analyst, and I will be facilitating today’s Interview. We are very pleased to have [Expert], who is an Engineer, Android Development at Guangdong Oppo Mobile Telecommunications Corp Ltd, with us today.

[Expert], before we start today’s Interview, please state if you agree or disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

Expert [00:00:32]: I agree.

Questions

[00:00:36] You have hands-on experience with many AI products in the industry. From a user perspective, since the release of GPT-6 Astra, what kind of changes have you seen in the number of tool-calling rounds and total execution time for long-horizon coding and terminal tasks compared to the previous generation? Have they decreased, and if so, by how much?

Expert [00:01:01]: I think this issue can be divided into two aspects: not just time consumption, but also the degree of task completion. There were many tasks that GPT-5.6 Sol simply could not complete, or performed so poorly that they were completely useless. Astra can complete many tasks that Sol previously could not, such as 3D modelling or, in some recently popular use cases, directly using computer-use capabilities to operate drawing software to create images, which are tasks that GPT-5.6 Sol was unable to do before. The second aspect is the time required. In my personal experience, for tasks like programming, if it takes one hour with Sol, it takes only 20 minutes with Astra, indicating a reduction of about 60%. However, OpenAI's own report states that the comparison results show a reduction of about 50%.

ES [00:02:13]: So your hands-on experience shows even more improvement than what its official statement claims; is that right?

Expert [00:02:20]: Yes, that's my personal experience.

ES [00:02:29]: What do you think of this improvement? Does its contribution come more from terminal command execution, or from computer use's help with graphical interfaces? Since you just mentioned that both directions have their own benefits, which aspect is relatively greater in magnitude?

Expert [00:02:51]: I think these two are not an either/or relationship. Since computer use also requires the execution of terminal commands, improving computer use capability must be predicated on improving terminal command execution capability. The model needs to understand what kind of content is on the screen, what the meanings of these contents and controls are, and which thing I want to click. Astra is much stronger than Sol in this regard. The time-consuming level I just mentioned does not include computer use, because the previous computer use was too difficult to use. Basically, I think there is no comparison. It can be said that it has made a leap forward. Therefore, the evaluation I just made is more about coding tasks.

[00:04:06] On the pricing front, we are seeing Astra cost about 2.5x more than the previous generation, with input at around USD 10 and output at around USD 50. From the perspective of the application-side cost structure, are clients willing to pay this premium? For which specific tasks would they justify the extra cost, and for which would they downgrade back to established, mature models?

Expert [00:04:30]: I think when the expected return of a task may be greater than its investment, clients are willing to pay for it. For example, many personal users may think they can make money by building something themselves with Codex. That is why OpenAI has been promoting the concept of OPC. Of course, this refers to expected returns, not actual returns. For a software company that is already profitable, with GPT-5.6, one engineer can do the work of two, and with Astra, one engineer can do the work of three. In that case, buying Astra is definitely cost-effective, even if that means choosing the top-tier plan at over USD 200.

Or consider a 3D modelling company that originally needed one expert and three junior staff members. With Astra as a tool, it may only need one expert and one junior staff member. The labour costs saved in this way are relatively small compared with the cost of Codex and ChatGPT memberships. As for API calls, I understand that most of them still use Chinese models rather than the top-tier model. Except for some high-value tasks in the financial sector that can generate enough revenue, API costs are relatively low. For example, in some low-cost workflows, low-cost APIs are used. As far as I know, DeepSeek is the main choice; for instance, DeepSeek can be integrated into AI customer service workflows.

One characteristic of this workflow is that it doesn't give the model much context, because it is split into nodes, and each node receives relatively little context. With limited context per node, even a switch to DeepSeek or another model won't degrade performance much; DeepSeek just happens to offer faster inference. This mirrors what Duolingo pointed out regarding its own app: it simply can't afford to run OpenAI or Anthropic models because the operational costs would be prohibitive.

[00:07:13] You can tell that Astra's computer use capabilities have been greatly upgraded. If we look at DeepSeek, now that it's gradually transitioning from V4-Flash-Vision to the native vision understanding of V4.1-Flash, where does it currently stand from an application standpoint, specifically regarding screenshot comprehension and UI element grounding success rates? How would you evaluate its performance?

Expert [00:07:40]: If I were to evaluate it, I can only draw an analogy, since I don’t have a very large dataset. If Astra is a junior high school student, DeepSeek V4 is still at the level of having just graduated from kindergarten, and it is completely unusable for understanding interfaces. This may also be related to the harness. For example, GPT is able to take screenshots of applications. However, if you use GPT as a relay to access DeepSeek's API, it cannot take screenshots, including when integrated into WorkBuddy and similar setups, so it can only recognise the entire page. In the case of the entire page, for example, if two windows are opened at the same time, it cannot understand that the two windows correspond to two applications. My personal speculation is that Chinese AI providers have not accumulated enough training data in this area, resulting in very poor performance and no usability at all.

[00:08:56] Among Chinese AI providers, which ones are relatively good at computer use? Is DeepSeek already doing quite well now? Is there still a gap compared with other vendors?

Expert [00:09:08]: Personally, I think DeepSeek is quite poor, and GLM-5.3 is also quite poor. In contrast, I tested the newly launched Doubao Work and found that, compared with harness frameworks such as WorkBuddy or DeepSeek harness, Doubao Work can actually complete some tasks. The level of task manipulation may be equivalent to that of GPT-5.6 Sol, but its speed is a bit slower. However, compared with DeepSeek, Doubao Work performs better in computer use.

[00:09:53] How do you view the pace at which Chinese AI providers are catching up in the field of computer use? Previously, Anthropic did relatively well in this regard; then DeepSeek caught up, while GPT has maintained relatively good performance, right? In your opinion, what is the expectation for Chinese AI providers to catch up? Will we see a better-performing model iterated this year?

Expert [00:10:21]: Which model will be better iterated this year? I can only say that there is still a lot of uncertainty in this regard, but I am personally optimistic about Moonshot AI. Moonshot AI is promoting a talent plan, which is equivalent to recruiting people in an unconventional way. I think the core of competition in the model field is talent, which depends on a company’s core values to attract talent. For example, Google has a lot of money, but many talents have left it. Although Alibaba has money, its model performance was ahead for a short period of time, but now it is [slower], and no one mentions it. At least Moonshot AI’s action makes me feel that it has a different understanding of talent.

As far as I am concerned, in terms of the model, judging from the values reflected by various companies at present, I believe DeepSeek will likely capture the majority of API call volume moving forward. DeepSeek has always designed the whole model architecture with cost as the core. Even if the price of APIs is very low, it can still make money from it, which is its biggest advantage. Second, when it comes to peak performance and top-tier capabilities, I think Moonshot AI's Kimi model might take the lead. For instance, with K3, even though its performance in other areas like long-horizon tasks falls short, and quotas for features like its coding plan are quite limited, its front-end capabilities have already reached the level of Anthropic

Combined with its emphasis on “taste” at the core of its AI philosophy in recruitment, I think it has the potential to compete with foreign models at the highest level. It is difficult to quantify taste. If it really finds a way to quantify taste, I think it will be a great competency. For example, although Astra has a strong ability to use computers and analyse code, it still lacks taste. For example, if Astra is asked to generate an article directly, short replies are fine; but if it is a full article, the output is very awkward and incoherent in every respect. Although the information is quite comprehensive, the way it is expressed is not fluent and feels awkward.

[00:13:30] MiniMax has historically focused heavily on the multimodal front, and it recently launched its H3 video generation model. Since computer use involves GUI operations, screen understanding and UI element grounding, do you think MiniMax will hold an edge and deliver strong performance down the road? After all, its M3 model hasn't seen an update since the beginning of the year, which is quite a while.

Expert [00:14:09]: I'm a bit out of the loop with MiniMax right now. Its iteration pace has slowed down for unknown reasons, and there hasn't been much news about it lately, so it's hard for me to give an assessment of the company right now. For example, DeepSeek V4-Flash was released at the end of July; roughly a month later, GLM-5.3 came out, and then about half a month after that, DeepSeek V4.1 was released. As I understand it, model iteration speed should be improving as expected. However, I don’t know what bottleneck MiniMax has encountered, so it’s hard for me to comment.

[00:15:00] Some while ago, DeepSeek adjusted its API pricing and released version 4.1, essentially routing most of the tasks originally handled by V4 Pro to the new model. While some clients still have the option to choose, the majority of user demand has already migrated to this more cost-effective alternative. In light of the impact of this price change on actual usage and the fact that some users chose to stop using DeepSeek during the price increase, if its subsequent pricing is on par with that of other model vendors, do you think it can still maintain its original usage advantage?

Expert [00:15:36]: I think its model has two advantages: besides being cheap, it also has a faster inference speed compared with other models. Therefore, even if the price increases to match the competition, DeepSeek remains the clear choice due to its superior speed. Furthermore, a minor price hike shouldn't significantly impact developers who have already integrated DeepSeek into their workflows. It is highly unlikely that a slight increase would break their margins or make the workflow unviable.

[00:16:18] Following the release of Z.ai’s GLM-5.3-Flash, have you noticed a significant diversion of API traffic away from DeepSeek Flash based on actual developer adoption? What is your rough estimate of the impact?

Expert [00:16:37]: As far as I know, the impact of GLM-5.3 on V4-Flash is not great because it has been slightly improved. Whether it is GLM-5.3 vs. V4-Flash or V4.1-Flash vs. GLM-5.3, their scores are a few points higher than the existing benchmark, without much change. As a developer, based on the actual business operation situation, it is impossible for me to change just because you have improved a little. This is because changing a model requires retesting many cases. If customer complaints are caused, it will be another risk. It is only worth rerunning these cases when there is a significant improvement, to compare the two models and see if switching is truly worthwhile.

[00:17:43] Beyond improvements in model capabilities, what are the primary migration costs for developers? Does it mostly come down to prompt refactoring, tool-calling compatibility or regression testing? How should we evaluate the scale of these costs, specifically, how do we quantify them in terms of time and budget?

Expert [00:18:06]: The core issue is testing. We need to run regression tests using the new model on test cases that have already passed. During this testing process, we might find that prompts, including system prompts and tool-calling instructions, need to be fine-tuned to improve success rates. So ultimately, it comes down to the cost of testing. In terms of actual impact, this cost is measured in engineering hours. For instance, an engineer might need to spend a whole week running through all these test cases just to evaluate whether switching models is actually worth it.

ES [00:18:50]: What kind of tasks roughly correspond to a week's workload? If these are B-end applications, wouldn't there be more factors to consider and a greater overall workload?

Expert [00:19:06]: What I just mentioned applies to the B2B side. The enterprise clients will accumulate some workflow execution cases, specifically, both passed and failed ones. When switching models, they need to run a test cycle. If they notice the success rate drops after the switch, they will definitely choose to either stick with the old model or fine-tune the prompts for the new one. So, the core issue here is ultimately the cost of testing.

[00:19:52] We talked about DeepSeek's capabilities earlier. You mentioned that it may still be at a kindergarten level when it comes to agent tasks, especially computer use. Where is the main room for improvement? And looking at current tasks, how are you tackling them? Are you leveraging overseas models to get them done?

Expert [00:20:20]: I almost exclusively use overseas models, but I may also use overseas models to build a workflow, somewhat like distillation. Distillation is about training a small model and using a good model to build a workflow with a slightly weaker model, reducing costs in scenarios with high token input/output. Some daily work definitely will not use Chinese AI providers' agents. Once you have used foreign agents, it is hard to go back to Chinese ones because there are too many areas that need rework. For example, give a task to Astra, and you can leave it alone for two hours, and it will have finished the job. But if it is run on a Chinese model, it may stop after 30 minutes, and the work it has done may still have many problems. This is completely unbearable. It is almost like having to watch it work with full attention to see where it has introduced a bug—this is completely unacceptable.

ES [00:21:41]: Do you see this more as a difference in capabilities at the model level, or as differences in dimensions such as the harness or agent framework at the architecture level?

Expert [00:21:53]: The problem lies within the model itself. Even if you proxy GPT or Codex to a Chinese model, it still performs poorly.

[00:22:14] What do you think of the gap between Chinese and overseas leading models? Some outsiders have quantified it. For example, the optimistic estimate is about 3-6 months, and some may think it will be around 1-2 years. According to your personal use of somatosensory, when will the current Chinese cutting-edge models be equivalent to overseas ones? What do you think of this gap?

Expert [00:22:55]: For example, I think DeepSeek may be at the lowest level of 5.5 Sol in long-range task processing, which is the last generation of GPT. It’s just that it is easier to use in workflows with a short context, but only if it reaches long context, it will completely collapse. The same is true for other models. For example, although GLM emphasises its processing capability for long-horizon tasks, its processing speed is very slow compared with GPT, and the products produced are relatively simple. In comparison, the recently launched Doubao 2.1 model may be relatively better in terms of computer use capability. However, for long-horizon tasks, I definitely wouldn't use Doubao either, because it has let me down too many times before, and its hallucination rate is too high.

[00:24:17] Now, the top-tier work agents in China include WorkBuddy; Lark Aily has also been integrated with Doubao, and Alibaba’s Wukong has also been integrated with DingTalk. What do you think of such products? In H2 of the year, major internet companies will likely all focus on promoting such products. Do you think any of them currently has a leading advantage in the Chinese domestic market? Which company might perform better? Is it Doubao? Nowadays, outside observers may mostly think that WorkBuddy has a relatively leading edge. What are the strategies and current progress of the top internet companies?

Expert [00:24:56]: I think that although WorkBuddy currently has a relatively high call volume, based on the several cases I have tried, Doubao Work has the potential to rise to second place in the Chinese agent field. Qwen tried its QoderWork before, and these things were very difficult to use. Therefore, I personally understand that Alibaba may fall behind in this regard. Judging from the current situation, I don’t see any good planning or possibility of breakthroughs. To me, what these domestic players are doing is basically mimicking OpenAI. Since OpenAI integrated Codex and ChatGPT, these guys are turning their own workplace tools, such as Lark and Tencent Docs, into connectors for agent apps, rather than just keeping them as standalone products. They're clearly trying to copy OpenAI's playbook. At the same time, the integration of work capabilities also means that the consumption of user tokens will increase, which is certainly helpful for their data collection and subscription user growth.

ES [00:26:34]: You think Doubao will grow faster in the future, but WorkBuddy will still account for the largest proportion. Is that right? Will Qwen and DingTalk fall behind?

Expert [00:26:52]: I don’t know much about DingTalk. As far as I know, no personal users will actively open DingTalk to use its AI tools. It should be the company that has more requirements.

ES [00:27:15]: These companies are all promoting the B-end market. When we talked about it earlier, they still charge for the B-end market. How do you rank the competencies of each company?

Expert [00:27:38]: First place still goes to WorkBuddy, followed by Doubao Work. As for Doubao Work, I'd say its advantage is just a short-term lead in its AI model. The biggest advantage of WorkBuddy is that the whole ecosystem is open. You can swap in your own models, use various plugins and leverage its Buddy apps. The customisability is just way higher. Therefore, I think WorkBuddy will continue to take the lead. It’s just that Doubao Work may catch up with it as a new product. Doubao Work does not support model switching for the time being. If a newer and more powerful model appears and it can't switch over, it will cause it to lose its competitive advantage instantly. Doubao itself has a large number of personal users, which is also one of its advantages. It’s just that it can’t get much money from individual customers because many people want to play with Doubao and ask for some emotional value.

ES [00:28:50]: Doubao has also started monetising. I think it rolled out a three-tier subscription model sometime in H1 this year. Rumour has it that their conversion rate is around 5%, with users paying dozens of RMB for memberships. In your opinion, is this primarily driven by lightweight work scenarios, or are there other needs at play?

Expert [00:29:17]: I understand that there may be some demand for video generation, and I don’t have a deep understanding of its application scenarios. As far as I know, the hallucination rate of Doubao is relatively high. It’s okay for personal users to ask it some simple questions. If you want to pay, it may involve this thing and he may make money. For example, they generate an AI video and post it on the Internet. They think that this video may become popular, so they pay some money to make a few videos. This is a scenario I personally speculate.

[00:29:56] OpenAI announced that it will stop providing models to Cursor starting 12 November. As a heavy user, where do you think these users might migrate? Will they move to Claude Code, Chinese providers' IDEs, self-built solutions, or OpenAI’s own upcoming integrated products?

Expert [00:30:24]: For anyone who has used GPT models, unless their company genuinely refuses to subscribe and they personally don't want to either, it is highly unlikely that they will switch to Chinese AI providers' agents. Because, as I just said, the reliability is too poor. Even after so many iterations, domestic models are still very unreliable for coding and have many problems. It's just that there has been a major improvement in a single area, namely front-end capability, such as the K3 just mentioned. The improvement in K3's front-end capability means that, for example, in my workflow, I need to generate some front-end pages—say, when making a PPT. The first few steps of a PPT don't require HTML, but the last step does. In that case, I can use K3 or GLM to generate some web pages, which gives a better result.

[00:31:30] For your own use case, how do the costs of tool subscriptions, direct API access, and self-hosting stack up against each other for the same volume? What’s the actual price gap? Also, do you think individual developers will shift more toward edge-deployed, distilled small models for mainstream tasks in the future?

Expert [00:31:53]: Personally, I think that is highly unlikely, primarily because edge deployment comes with its own hidden costs. While it can save money on paper, we have to ask: what kind of independent developer would actually choose this route? It would likely be someone who is perfectly fine tolerating the performance limitations of Chinese models and is willing to buy dedicated hardware for it. Alternatively, it appeals to those who simply don't know how to access superior overseas models. For instance, some vendors sell hardware bundles pre-loaded with OpenClaw-type software, marketing them essentially as plug-and-play "digital employees", and pitch them to buyers who can't tell the difference between Chinese and top-tier foreign models. If a buyer doesn't know what peak performance looks like, edge deployment works. Otherwise, for savvy developers, the time cost of building, configuring and troubleshooting their own hardware rig is simply way too high to justify.

ES [00:33:15]: Under the same usage volume, what is the cost difference between buying usage directly through a tool subscription and connecting to its API?

Expert [00:33:30]: The difference is significant. I once saw a statistic that with GPT at USD 200, plus its constant resets, you can get up to USD 2,000 worth of usage. In that case, USD 200 is still very cost-effective. Although it has now launched a plan for recharging usage after the subscription ends, I am not sure whether others will choose to buy usage this way. If it were me, I would rather buy a few more accounts and switch between them than buy its usage, because it is not cost-effective.

[00:34:12] What are your thoughts on Meta’s recent launch of Muse? It’s seen strong initial traction in the US in terms of downloads and user feedback, but how do you view its long-term sustainability? Meta seemed to fall behind for a while. What’s your take on this new strategy and its current execution?

Expert [00:34:46]: I haven’t heard of this.

ES [00:34:50]: It is a personal product, a personal AI assistant, or a personal agent.

Expert [00:35:02]: I haven’t heard of this product, so it is hard for me to evaluate it.

ES [00:35:19]: What is Meta’s current action at the model level? Hasn’t there been much performance recently?

Expert [00:35:31]: Meta's recent models haven't really blown anyone away. It did drop a new model, but the reviews have been lukewarm at best. It's nowhere near the level of Astra; honestly, it feels like it's just patting itself on the back. Since it already ruined its reputation by benchmark-gaming in the past, it seems like fewer people are even bothering to test its new models now.

[00:36:12] What do you think of the capability gap between Chinese and overseas models in the next year? Now, frontier overseas model providers have also mentioned that they will slow down or pace the R&D of models. Will this give Chinese AI providers some opportunities to catch up? Or will this gap not be narrowed?

Expert [00:36:29]: I think the slowdown has become a means of hype. They talk about slowing down, but in fact, they invest more and more. Therefore, this speed reduction is just to make people who haven’t paid attention to AI pay more attention to it. The bottleneck lies in computing power. For example, Astra’s 20X package is not available because its computing power cannot keep up with it. Compared with GPT-5.6 Sol, the usage of Astra is faster or smaller.

As they hit a compute bottleneck, it creates a window of opportunity for Chinese models to catch up. For instance, even if they develop a brilliant AI iteration process within the next six months that automates most of the research work, it won't matter if they don't have the compute to actually run or train the new models. If their next-gen models require massive compute that they simply cannot access for training or inference, China could leverage this downtime to close the gap.

Computer use capabilities, I think China’s models will definitely catch up within the next six months. As long as the input and output of GPT models are obtained, these data can be distilled. Distillation is also correct or used as a benchmark for evaluating models, which can make up for computer use capabilities. The improvement of computer use ability is also helpful for the improvement of long-range tasks, because many long-range tasks do not involve direct manipulation but visual recognition. For example, if someone wants to develop a relatively large front-end website, this website needs screenshots to understand whether there is any problem with the layout of this website. Computer use improves the ability of agents to perform long-range tasks, or so it is said.

ES [00:39:15]: Should the upgrade of computer use also bring about a clear increase in token usage?

Expert [00:39:29]: I think the increase in computer usage will increase in the short term. Everyone has come up with many amazing cases, but in the end they found that these cases can only be regarded as toys for people who don’t know what their use is and will not bring much improvement. If I have a 3D printer and a child is given something fun every day, I can make it with Astra. This kind of scenario is okay. If others just see that someone else has run a 3D model on the Internet and try it by themselves, but after running several models, they find that there is no interest and nothing will be done. I think this is just a short-term improvement.

[00:40:22] We mentioned the work agents of leading internet companies earlier. Should we pay more attention to the computer-use capability of basic models used in such products? Which one will improve faster? Is this part of the improvement helpful for actual work scenarios?

Expert [00:40:44]: As far as I am concerned, even though Astra can run now, there are still some critical points that need to be crossed between running and usability in terms of speed, cost and accuracy. If DeepSeek can reach this level of speed in text processing and reasoning, and its cost is as low as that of DeepSeek, and its click-through, execution accuracy and rework rate are relatively low, it will be possible only after crossing such a critical point. Of course, it is hard for me to quantify this. For example, if Astra currently takes 10 seconds to perform one click, and that could be reduced to two seconds, then I think computer use could become practical. Because even now, using Astra for some computer-use operations is still too slow. For people who are already skilled at using software, it hasn't crossed the critical speed threshold.

ES [00:42:01]: Is this mainly a computing power bottleneck, or is it a problem at the model level?

Expert [00:42:10]: Why is Astra's inference speed slow? I am not a professional model researcher, so I am not in a good position to evaluate that.

ES [00:42:24]: However, if computing power improves further, can this click speed be accelerated further? Or is it a limitation of the model itself?

Expert [00:42:33]: I think this is a problem with the model itself, because if it were a computing power issue, there should be a fastest case. But in fact, there is no fastest case either, so it is most likely a problem with some of the model's internal reasoning algorithms.

ES [00:42:55]: This expectation should improve significantly next year, right? Because computer use is no longer a new feature, GUI-based agents had already begun to be released in early 2025, with some use cases, but their usability was very poor at that time. After a year and a half to two years, we have seen a clear improvement. By H2 2027, do you think this area will produce good results?

Expert [00:43:26]: I think this critical point will definitely be crossed by this time next year, and it might even be crossed within half a year, because this is a relatively well-defined task for AGI to complete. Nowadays, people have quite different definitions of AGI. For example, Jensen Huang currently says AGI has already arrived, and there are various other views. Personally, I think AGI has two characteristics: first, it can iterate on itself; second, through self-iteration, it can gradually replace all professional work performed by humans. What is the precondition? It needs sufficient data. If we want to discuss subsequent model development, I think the most important problem lies at the data level. For example, whether it is computer use or not, why hasn’t this area been developed in China? I think it is a problem at the data level.

Since no one in China is paying attention to high-quality data collection, we can only use the output of foreign models as training data. For example, Z.ai may be weighing the pros and cons. It also launched a promotion before that it would give some professionals a membership card for one year, with a value of several hundred RMB. Many people complain that the daily salary of real professionals may be more than this. It is impossible for you to deceive others with such a price and contribute professional data. I think the reason why it makes such a trade-off is that it takes into account the investment in manpower and the higher cost of using other people’s ready-made data. Therefore, it has no great motivation to accumulate data in this area. However, I think the only way for Chinese models to surpass overseas models in the future lies in their data collection capabilities and having a data volume that reaches or even exceeds that of overseas models. Otherwise, they will always just be training on other people's data.

[00:46:28] A couple of days ago, Anthropic also released a report on model distillation, naming several Chinese companies. In your observation, what specific areas have become more difficult for leading Chinese model developers when distilling some overseas models? Going forward, will there be less room for distillation, or will there always be some workaround?

Expert [00:46:52]: I can’t really speak on that too much, since it falls into a bit of a grey market area. They have their own workarounds, and someone will always figure out a way to get the data. For instance, if I set up a shell company overseas and have people who don't even use computers register accounts for me, wouldn't that be pretty straightforward? So, I think there will always be a workaround. It’s just that for regular users based in China, getting access to a Claude account is going to become much harder.

[00:47:36] With native frameworks transitioning to open-source, what's your take on the trajectory of single-task token consumption? What kind of growth magnitude should we expect, and how much value do these open-source frameworks actually bring to the ecosystem?

Expert [00:48:17]: I don’t think the per-task usage has much to do with whether the original framework is open-source or not. Even if there are no DeepSeek harnesses or Codex harnesses, single-task API calls would definitely keep climbing anyway. That’s because AI penetration is still relatively low right now and is bound to keep growing. It’s just that after the wiring harness framework comes out, there is more room for everyone to struggle. However, most people won’t do it. There are only a small number of people who sell their own good agents. In other words, there is no need to design an entire agent application from scratch.

In addition, I noticed that you mentioned LangChain in your previous question. As far as I know, the competition between LangChain and DeepSeek harness is not in the same dimension. For example, DeepSeek harness includes various extra features such as conversation and skill retrieval. However, if the agent is not presented entirely as a chatbot, I don't need these extra features at all; they become a burden. In that case, LangChain is actually a better choice. For example, if I want to present the interaction with the agent to users in a dialogue box or as a chatbot, then features like conversation are useless. It is sufficient to directly add some prompts and tools with LangChain; there is no need for UI elements, and the extra functionality becomes a burden.

[00:50:10] Finally, do you have any personal insights on this topic that you could share to supplement and summarise for us? Based on your observation of current usage scenarios or applications, do you think there are any key opportunities worth watching in H2 this year and next year, opportunities with relatively large growth potential or clear trends in the AI industry as a whole?

Expert [00:50:43]: Personally, I am quite optimistic about these data collection companies, whether Chinese or overseas ones. In my opinion, their position is similar to that of CATL relative to automakers: many carmakers are losing money, while CATL is profitable, and all carmakers rely on CATL. The same is true for data. For example, foreign companies such as Scale AI and Surge AI are collecting data, but their valuations are only up to USD 29bn, while Anthropic and OpenAI are aiming at trillion-level valuations. I think they still have huge room for growth. As for domestic companies, I haven’t heard of any that do particularly well in data collection, but if there are any, I think they would be excellent.

Another direction is on the learning front. For example, Andrew Ng is working on a learning platform—the startup project he is currently working on. I think this is also a promising direction. After using AI, it becomes clear that one’s motivation to collect information, reason, think, and learn will all weaken. In this context, whoever still has the ability to think and learn independently will be the one to stand out in the future. When I talk about standing out, here's the situation: everyone is using AI, and AI is stronger than ordinary people. How can one stand out when the average level has been elevated? At this point, we need to consider whether a person's abilities can surpass AI. The prerequisite for surpassing AI is that the person must continue learning and thinking. So I think learning will be a very important direction.

[00:53:10] In terms of downstream applications, the two areas with the most successful commercial traction right now are coding and video generation. Given that the market is also closely watching AI for science, have you observed any other use cases that are gaining momentum and have the potential to catch up, becoming the next major, clearly defined commercial market?

Expert [00:53:51]: Let's start with the obvious: the promise in fields like finance, law and medicine is all apparent to everyone. Even Kimi has been highlighting its capabilities in handling financial and research tasks, so finance should be a promising area. Aside from these, we haven't really seen any industry-wide, general-purpose products emerge yet; I haven't seen any so far. It is still more about customised solutions, such as FDE, which is very popular right now. I don't think FDE companies belong to any specific industry, but they can empower companies across all industries. Therefore, FDE service providers are worth paying attention to.

Since many industries use their own business systems, general-purpose agents cannot access them. To bridge the tooling gap, there need to be people who understand how to provide the data AI requires. Therefore, I think FDE will be a major growth point. In the current market, I have observed that many people want to do FDE, but it is difficult to gain clients' trust. If there is a leading FDE company, things may move faster.

[00:55:46] Z.ai used to focus heavily on the B2B segment, handling FDE [forward deployed engineer] tasks entirely in-house. Do you think traditional IT outsourcing firms like Chinasoft and iSoftStone can successfully pivot into top-tier FDE companies in the AI era? Or is the entry barrier too high, meaning their legacy expertise in business systems and client scenarios won't easily translate to AI?

Expert [00:56:12]: I think it will be quite difficult. First, I haven’t heard of any notable AI transformation initiatives from them recently. Second, they rely on manpower outsourcing. Whether they have the ability to cultivate and recruit such talent is another question. If it is manpower outsourcing, one issue is whether they have the ability to train and recruit. The second is project outsourcing, but this model does not hold in the FDE field, because project outsourcing requires the client to provide clear requirements, while FDE is about figuring out the requirements. So the two are completely different. I’m not very optimistic about them for now; instead, I favour the AI service providers that were founded in the AI era.

ES [00:57:07]: Which domestic companies are doing relatively well and making rapid progress right now?

Expert [00:57:21]: In China, the one I have seen that feels relatively professional is Datawhale. Datawhale has a WeChat official account; you can follow it. Tencent and Alibaba are also promoting FDE, but they tend to select larger-scale enterprises, so they cannot completely take over this market.

ES [00:57:56]: The FDE business is inherently impossible to standardise; it is somewhat similar to the previous customised project business.

Expert [00:58:07]: Well, nurturing people is very important, as well as whether they can see the commonalities among different problems across different industries.

ES [00:58:21]: By how much do you think the gross profit margin of FDE projects is higher than that of traditional project outsourcing? What are their gross profit margins?

Expert [00:58:43]: I don't have a deep understanding of the gross margins in this industry. However, based on a scenario I’m familiar with, say, a company wants to drive an AI transformation across its product lines and internal roles; they might invest around RMB 200,000. For that budget, an FDE service provider might assign two people. If their combined monthly salary is RMB 80,000 (RMB 40,000 each), the gross margin indeed looks quite tight. Since I can't give you an accurate assessment, I'd rather not speculate. You might need to consult industry insiders to get the real picture

ES [00:59:33]: Thanks again to [Expert] for their input today. This now concludes our meeting. Goodbye.