The junior engineer no longer types the boilerplate by hand. She describes what the function should do, and the editor writes the first draft. The company's own internal metrics now count how much of every commit the tool wrote.
A year ago, AI coding tools in China were a curiosity developers tried on the weekend. In 2025 they became part of the workday. Three names now define the local market: Trae from ByteDance, Comate (文心快码) from Baidu, and Tongyi Lingma (通义灵码) from Alibaba Cloud.
Three tools, one shift
The pattern across all three is the same: an AI (人工智能) layer that writes, explains, and fixes code inside the editor, not in a separate chat window.
- Trae reported, in its 2025 annual product report, over 6 million registered users and 1.6 million monthly active users by December 2025, with roughly 100 billion lines of code generated during the year.
- Comate (文心快码) said at the end of 2025 it had served more than 8 million developers and over 2,000 enterprise clients, with an average code-adoption rate around 38%.
- Tongyi Lingma (通义灵码) passed 22 million plugin downloads by September 2025 (up from 15 million in May) and had generated over 6 billion lines of code for developers.
These are vendor-reported numbers, but the direction is consistent across all three: the tools are no longer being tried; they are being used.
Trae: free, and Chinese-first
ByteDance launched Trae in January 2025 as, by its own description, China's first Chinese-language AI IDE (integrated development environment). The strategy was to give it away: the personal version is free, with no usage cap, and it bundles access to strong models including DeepSeek and ByteDance's own Doubao (豆包).
The most telling number is internal. At ByteDance's Volcano Engine conference in mid-2025, a company technology vice president said over 80% of ByteDance engineers were already using Trae, and that the firm had mandated it internally, replacing tools like Cursor and Windsurf. The open-source core, Trae-Agent, was released on GitHub in July 2025 and has since gathered roughly 11,800 stars — a rare case of a Chinese coding tool contributing code back to the global developer community.
Comate: the enterprise floor
Baidu's Comate (文心快码) leans toward the workplace rather than the individual. Built on the ERNIE (文心) model, it plugs into more than ten mainstream IDEs and handles over 100 languages. The adoption figures are the interesting part:
- At Himalaya (喜马拉雅), a reported 44% of generated code was adopted.
- Inside Baidu itself, the share of daily new code written with Comate passed 40%.
- Across enterprise users, Baidu measured an average 38% adoption rate — meaning more than a third of suggested code survives into the product.
For a large company, that ratio is the business case: a tool that writes two-fifths of the routine code changes the staffing math without changing who owns the system.
Lingma: the download leader
Alibaba Cloud's Tongyi Lingma (通义灵码), launched in late 2023, has the widest reach by installs. Beyond the 22 million downloads, it became the only Chinese product in Gartner's AI code-assistant Magic Quadrant as a challenger in 2024 — a third-party benchmark, not a self-award.
Internally at Alibaba Cloud, the share of code written with AI tools reached roughly 40% by mid-2025. Adoption shows up in named customers: China Construction Bank's tech arm, Ping An, Geely, NIO (蔚来) and XPeng (小鹏) all report double-digit to majority AI-generated code in parts of their stacks. At NIO, the AI-code share runs above 30%; at XPeng, code-review acceptance of Lingma's suggestions reached 50%.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the Chinese AI-coding market is moving from "autocomplete toy" to "agentic teammate" faster than most Western analysts assumed, because domestic tools were built Chinese-first and priced to spread.
What "adoption" really means
The honest read is narrower than the headlines. The productivity gains vendors cite land in the 10%–30% range for most teams, not the "10x engineer" myth. Developers trust AI most for the two tasks that dominate real work: bug fixes (about 40% of Trae sessions) and code generation (about 30%). They stay cautious on repository management, environment setup, and optimization — exactly the parts where a wrong move is expensive.
The tools also did not erase the need for engineers. As one ByteDance executive put it, even if 85% of the code is written by AI, a human still drives the process. The scarce skill is shifting from typing syntax to specifying intent and verifying the result.
Honest limitations
- All user and line-count figures (Trae's 1.6M MAU and 100B lines; Comate's 8M developers and 38% adoption; Lingma's 22M downloads and 6B lines) are vendor-reported, not audited by an independent body, and "lines generated" counts suggestions, not shipped, code.
- The market-size claim of about RMB 4.5 billion (≈ US$630 million / HK$4.9 billion) for China's AI-coding tools in 2025, up 87% year on year, comes from industry recaps citing IDC; treat it as an estimate.
- Reported adoption rates (Comate 38–44%, Alibaba Cloud ~40%, customer cases 30–50%) depend on how each firm defines "adopted" or "AI-generated," and these definitions are not standardized.
- Gartner's 2024 Magic Quadrant placement is a 2024 snapshot; the field has moved since, and rank is not a durability guarantee.
- Security and compliance remain open questions for enterprises handling sensitive code; most vendors offer private deployment, but adoption of those options varies.
What readers can do now
- If you write code, install one of the three and use it on a real task this week — a bug fix or a test scaffold — then check how much of its output you actually kept.
- If you lead a team, measure adoption by accepted lines, not installs. A tool nobody merges is a sunk cost; a tool at 30% accepted code is changing your throughput.
- Watch the enterprise numbers, not the consumer hype. Comate's 2,000-plus enterprise clients and Lingma's bank and auto customers are the signal that this category has crossed from toy to infrastructure.