Alibaba, ByteDance, and Tencent are collectively making heavy investments in new emerging sectors.

In the second half of this year, AI Coding has become the hottest track in the industry and has taken the lead in opening up the prospects for the commercialization of AI.

“From Silicon Valley to China, everyone says they're working on Coding, and there are too many projects to review,” a technology investment professional told China Entrepreneur.

In September this year, leveraging its leading position in the Coding field, Anthropic completed a Series F financing round of $13 billion, with a valuation reaching $183 billion. This valuation has tripled compared to just six months ago.

The application value of Coding has enabled Anthropic to compete with OpenAI, which has 800 million global users, in the B2B market. With a 32% market share, leading OpenAI's 25%, it has become the most commonly used model provider among enterprise users. Its revenue soared from $1 billion in 2024 to $4.5 billion in the first half of 2025.

The fire of Coding R & D is also burning brightly in China. Both large companies and startups launched independent IDE (Integrated Development Environment) products in 2025.

In March this year, ByteDance released the domestic version of its Coding tool, Trae. By May, the monthly active users of this product exceeded one million. In July, Tencent initiated the internal testing of its AI Coding assistant, “CodeBuddy IDE”. In August this year, Alibaba launched its AI Coding platform, Qoder.

Star startups in the large - model field are also not to be outdone. In July this year, Darkside of the Moon released the Kimi K2 model, with significant improvements in programming, Agent, and long - text functions.

In August this year, DeepSeek's DeepSeek - V3.1 model integrated with the Anthropic ecosystem, allowing users to easily integrate the capabilities of DeepSeek - V3.1 into the Claude Code framework.

The external environment is also intensifying the urgency of large companies' Coding R & D. On September 5th, for various reasons, Anthropic suddenly announced that it would stop providing services to Chinese companies. Next, which large company or startup has the greatest hope of becoming China's Anthropic and getting the first first - class ticket in the AI field?

Why is Coding so popular, and what exactly is it?

Why are all the giants targeting the AI Coding track? Recently, Wu Yongming, the CEO of Alibaba Group, gave a clear answer. In the middle of September, during a speech at the Yunqi Conference, Wu Yongming stated that natural language is the programming language in the AI era, and Agent is the new software. “Developing the Coding ability of large models is the inevitable path to AGI.”

“Currently, Agents are still in the early stage, mainly solving standardized and short - cycle tasks. To enable Agents to solve more complex and long - cycle tasks, the most crucial factor is the Coding ability of large models. Since Agents can code autonomously, in theory, they can solve infinitely complex problems, understand complex requirements like an engineering team, and complete coding and testing independently.”

So, what exactly is AI Coding?

Simply put, it uses AI to assist users in programming, deeply integrating artificial intelligence technology into the entire software development process to lower the threshold and improve efficiency. With Coding tools, ordinary people can develop websites and mini - programs using natural language, and programmers can use them to complete and generate code. In the IDE integrated environment, Coding tools can also assist in the entire process of product design, testing, and synchronous modification.

For large companies, improving code production efficiency means lower costs and higher profits. Currently, large technology companies have entrusted a significant proportion of programming work to AI. Cai Chongxin recently revealed that 30% of Alibaba's current code is generated by AI.

For small companies and users, using Coding tools for product development allows ordinary people to cross technological barriers and achieve equality. At the same time, they are also the target customers that large basic - model companies hope to serve.

The AI programming ability is both a natural result of the enhancement of model capabilities and will help the model capabilities evolve and iterate. Therefore, the Coding ability can largely represent the strength of large companies' large models, and the commercialization path has been verified. In the future, for companies that want to build a good platform ecosystem and enrich Agent applications, doing well in Coding will be an inevitable task.

Dario Amodei, the founder of Anthropic, said in an interview that the number of programming users is growing very fast. “As the model becomes stronger in programming, it can also help us train the next stronger model. This is a very advantageous positive cycle.”

Competition among large companies: who takes the lead?

Actually, most large technology companies spotted the opportunities in artificial - intelligence coding a long time ago.

Recently, Yao Shunyu, a former researcher at OpenAI, referred to Coding as the main R & D task during an interview with “Language is the World”. He predicted that all large companies will improve the Coding ability of their models, and all pre - training, post - training, and reinforcement learning will take this into account.

Take Tencent's CodeBuddy as an example. Tencent revealed that currently, over 90% of Tencent's in - house engineers are using CodeBuddy, and the overall coding time has been reduced by an average of over 40%. In newly added code, the proportion of AI - generated code exceeds 50%. Combined with large - scale in - house production experience, R & D efficiency has been improved by over 16%.

The product form of CodeBuddy germinated in 2022. It has gone through three development stages: a plug - in, the CodeBuddy IDE form, and CLI (the product is named CodeBuddy Code). Currently, these three forms coexist, allowing users and enterprises to choose according to their needs.

Wang Shengjie, the person - in - charge of Tencent Cloud's developer AI products, told China Entrepreneur that from 2021 to 2022, some developers within Tencent proposed using AI to quickly understand code documents and assist in developing repetitive business code. “The slogan at that time was 'Tab Tab Tab No backspace', hoping that AI could assist in generating application code once.”

In the second stage, some business teams at Tencent proposed that simply generating code snippets was not enough. They hoped that the code tool could understand engineering projects and even have functions like unit testing. So, Tencent launched the Chat and Craft modes of Tencent AI Code Assistant (later renamed Tencent Code Assistant CodeBuddy) internally, with humans in the lead and AI in an auxiliary role.

As R & D deepened, Wang Shengjie's team gradually collaborated with the Hunyuan large - model team to develop a plug - in and integrate it into IDEs such as VS Code, which are frequently used by developers, to help developers with code completion, annotation, and recommendation.

In 2025, with the emergence of intelligent agents and multi - agent collaboration, Wang Shengjie believes that true AI Coding has begun and will soon undergo revolutionary changes. Coding will penetrate the entire product development lifecycle, from code generation to product internal testing, feedback, and adjustment.

Not only Tencent, but ByteDance has also been actively involved in Coding development. Hong Dingkun, the vice - president of technology at ByteDance, said in May this year that there are three reasons for ByteDance to seriously engage in AI Coding: to help more people master code to perform more complex tasks, to improve the work efficiency of professional engineers, and to help the model pursue a better intelligence ceiling.

ByteDance revealed that as of June, three months after the launch of Trae, its monthly active users had exceeded one million. Within ByteDance, over 80% of engineers are using Trae for auxiliary development.

Alibaba, which has made an all - out effort in the large - model field, is more committed to the Coding track. On the one hand, Alibaba continuously improves the coding ability of its Qwen basic large model and has also launched a dedicated Coding model. At the same time, Alibaba has introduced an independent Coding terminal product, Qoder.

At the recent Yunqi Conference, Alibaba released seven model updates, two of which directly enhanced the upper limit of Coding ability. Qwen3 - Max has a total parameter count of over one trillion, highlighting Coding programming ability and Agent tool - calling ability. The Qwen3 - Coder intelligent programming model once became the second - most popular Coder model globally on the Open Router platform, second only to Claude Sonnet 4.

On the independent product front, Alibaba's Qoder also has two major features. First, while ByteDance's Trae and Tencent's CodeBuddy both offer model selection, Qoder does not. The technical staff at Qoder explained that machine selection is better and faster than manual selection. They hope developers can compare the results from the perspective of achieving high - speed, high - quality, and cost - effective outcomes.

Qoder also has a higher pricing. For global users, the subscription fee for Pro users is $20 per month, and for Pro+ users, it is $60 per month, which is on par with the monthly fee of Cursor, an overseas leading AI Coding company. In comparison, the international version of Trae costs $3 in the first month and $10 per month thereafter. The domestic version of CodeBuddy is currently free, but using the Pro version requires accumulating points. Second,

Compared with Alibaba, ByteDance's Trae and Tencent's CodeBuddy have not developed dedicated Coding models and rely on ByteDance's Doubao and Tencent's Hunyuan large models.

Regarding the impact of having a dedicated large model, Wang Shengjie said, “Model capabilities will affect the calling and generation effects, but currently, the corpora of large models all have some code - related capabilities. We need to focus on balancing performance, quality, security, cost, and other aspects from a product perspective. We collectively refer to these as the product experience.”

Tencent's CodeBuddy still has its advantages, namely the Tencent ecosystem and enterprise - level applications. Wang Shengjie said that Tencent will not completely follow Cursor but will create differentiation. For example, it can connect to Tencent Cloud's assets, including WeChat mini - programs, enable one - click deployment of applications to the cloud, and carry out deployments for enterprise - level applications.

What will be the focus in the next stage: product experience and context engineering

While Coding helps improve model capabilities, the competition in this field is accelerating. At the end of September this year, Anthropic released Claude Sonnet 4.5. The new model can program continuously for 30 hours, generate 11,000 lines of code in a single instance, and even reconstruct an entire codebase.

The enthusiasm for financing in the Coding field is also rising. Foreign media reported that Cursor is in talks for a financing round of at least $1 billion, with a pre - financing valuation of $27 billion, which has tripled compared to its valuation three months ago.

From the start of its public beta in March 2022 to the end of 2023, the ARR (Annual Recurring Revenue) created by Cursor's four - person team exceeded $1 million, and the daily active users exceeded 30,000. As of June this year, its ARR had exceeded $500 million. The media predicts that this figure may double by the end of the year.

It's worth noting that Cursor does not develop its own large model but uses external large models such as GPT and Claude through APIs, focusing on optimizing the product experience. This shows that although Coding is closely related to large - model capabilities, the outcome of the competition does not solely depend on the model itself. Product experience and understanding of users are the key factors.

For example, Anthropic's product, Claude, has become the first choice for technical personnel. Apart from its high reliability and low hallucination rate determined by the model, it has also made a lot of detailed optimizations for developers. For instance, the code format is more readable, the interaction interface is user - friendly, the tool integration is rich, and it has a deeper understanding of edge scenarios. In addition, Claude has done a better job in data isolation and privacy protection in model design, which also meets the security requirements of enterprises. All these are inseparable from Claude's in - depth exploration of the Coding scenario.

Wang Shengjie said that the logic of an AI Coding product is to be more efficient, provide a better experience, and generate code quickly and accurately, and optimize the agentic workflow.

How to achieve speed and accuracy? Wang Shengjie believes that the current technical key lies in context engineering. In the programming scenario, developers often need to handle multi - file projects, long code blocks, or complex business logic, which requires the model to “remember” and correlate a large amount of context information.

Improving context capabilities requires enterprises and users to refine the logic together. “To build a good Coding infrastructure, users need to input higher - quality context memory content and generate Wiki (document) descriptions, including not only the code itself but also auxiliary content related to the code,” Wang Shengjie said.

Actually, Claude's leadership also lies in its ability to support an ultra - large - scale context window of up to one million Tokens. This enables Claude to fully “digest” the entire project's code structure, documentation, and even historical conversations. The model can generate more coherent code that conforms to the overall project logic based on global information.

Context has also become a frequently mentioned term by Alibaba's Qoder team. Currently, Qoder can support a context length of 200K, which still lags behind Claude. A technical staff member of the team said in a speech, “In the past, we thought 128K was enough, but now we find that even 200K and 300K are not sufficient.”

In addition, cost, efficiency, and accuracy are forming an “impossible triangle” for Coding products. After the launch of Qoder, many users complained about its rapid Token consumption, which is also one of the challenges faced by all current AI Coding products.

Wang Shengjie said that Tencent will analyze product data within its internal teams and finally optimize the thinking process of AI. “We will evaluate whether there is room for optimization in each round of calls, such as whether there is duplicate content and whether it is possible to try a more suitable model, so as to improve overall efficiency.”

However, the biggest challenge for large - model technicians may still lie in the extremely rapid generational changes in technological evolution.

A technical staff member at Qoder sighed in a speech: He has never seen a software engineering project or an efficiency - enhancing product develop at such a rapid pace. 'All paradigms only have a one - year lifecycle.'

Taking code retrieval as an example, from 2023 to 2024, the team where this technical staff member works used the traditional RAG (Retrieval - Augmented Generation) mode for code semantic retrieval. In 2025, it has shifted to a context - based retrieval method, which comprehensively uses semantic retrieval engines, keyword retrieval engines, code graph engines, and architecture knowledge retrieval engines through the Wiki approach.

As technologies such as Coding and context continue to break through, the spring of Agents may truly arrive. Recently, Yang Zhilin, the founder of Darkside of the Moon, said in an interview with “Language is the World” that Coding Agents are an important subset of tasks. “Ultimately, we hope to do more than just Coding. Even the models we are currently training are not just for Coding because it has some limitations.”

Currently, it seems that Agents and Coding are like fire and oil, and they may produce a more intense chemical reaction in the next few months.

This article is from the WeChat official account “China Entrepreneur Magazine” (ID: iceo - com - cn), written by Yan Junwen and published by 36Kr with authorization.