The Great Shake-up of LLM Open Source 2.0: 60 Out, 39 In. AI Coding Goes Crazy, and TensorFlow Is Dead

The guide for over - achievers has been updated again, and there is an extra episode this time.

Open Source 2.0: A Transformation as Dramatic as Plastic Surgery

After waiting for more than a hundred days, the suspense was finally revealed.

On the morning of September 13th, the open - source team of Ant Group ("Open - Source Technology Growth") unveiled the 2.0 version of "The Panoramic Map of the Large - Model Open - Source Development Ecosystem in 2025" at the Bund Summit in Shanghai.

Three months ago, the assertion of "A Real - world Hackathon in a live stream" still holds true today —

At that time, the "tear - off sheet" recorded the initial appearance of the ecosystem, and now, it has changed drastically.

Access address: https://antoss-landscape.my.canva.site/

This time, the panoramic map includes 114 projects (21 fewer than the previous version), covering 22 fields. Among them, 39 are new projects, and 60 projects have disappeared from the stage, including some once - shining Star champions —

For example, NextChat, OpenManus, FastGPT, and GPT4All were surpassed by later - comers due to slow iteration and weak community support.

The most dramatic event was the exit of TensorFlow. This former superstar finally couldn't withstand the attack of PyTorch, and the latter has since dominated the market.

The gray part represents the open - source projects that have been eliminated.

The overall trend is obvious: the ecosystem is undergoing a drastic reshuffle. Just like the "Cambrian explosion" of life, the Agent layer is the most turbulent, and in the chaos, various new species are emerging in an endless stream.

Another set of data also confirms this vigorous metabolism from the side —

Counting the eliminated projects, the "median age" of the entire large - model ecosystem is only 30 months, and the average lifespan is less than three years. It is an extremely young jungle.

Especially after the "GPT Moment" (October 2022), 62% of the projects were born, and 12 of them are new faces in 2025. That is to say, new projects appear and old ones fade out almost every quarter.

Even more exaggerated is that these young projects have received unprecedented attention: the average number of Stars is close to 30,000, far exceeding that of open - source projects of the same age in the past. The top ten projects almost cover the entire chain of the model ecosystem and are the most representative community forces at present.

Top 10 Most Active Open - Source Projects

The keyword word cloud also echoes this trend: AI, LLM, Agent, and Data have become the largest and brightest words.

The keywords AI, LLM, Agent, Data, and Learning in the large - model development ecosystem are the main fields of the projects listed in the first chart.

Another major change is the classification framework of the panoramic map.

Since I had seen the 1.0 version, when I first saw the 2.0 panoramic map, the most intuitive feeling was that the classification architecture had become more specific and detailed:

It has evolved from the general Infrastructure / Application to three major sectors: AI Agent / AI Infra / AI Data, clearly outlining the industry hotspots (centered around intelligent agents) and the trend of technological evolution.

If the 1.0 framework still bore the shadow of the traditional open - source software ecosystem, then the 2.0 version already shows the characteristics of the "intelligent agent era".

Finally, from the perspective of the profiles of 366,521 developers globally, China and the United States, as the two leading countries, contribute more than 55% and are still the absolute leaders of the projects. Among them, the United States ranks first with a proportion of 37.41%.

In the detailed contributions in the technical field, the United States has obvious advantages in AI Infra and AI Data.

For example, in AI Infra, the United States' contribution rate reaches 43.39%, more than twice that of China, which ranks second; and its leading advantage in AI Data is even more obvious.

China performs similarly to the United States in the specific application layer (AI Agent). The contribution rates of the two countries are 21.5% and 24.62% respectively, which is closely related to the greater investment of Chinese developers in the Agent layer.

The Evolution of the Mapping Theory

Why put the methodology in the front? The answer is simple — which projects can enter the 2.0 panoramic map largely depends on the change in the methodology.

The methodology of the 1.0 version was "starting from the known" — the widely - discussed leading projects, such as PyTorch, vLLM, and LangChain, and then extending outward through their collaboration and dependency relationships.

But the starting point determines the boundary: which seed projects you start from determines the scope of the ecosystem you can see. At that time, the threshold for selection was an average monthly OpenRank (an open - source influence indicator developed by X - lab of East China Normal University) of ≥ 10.

For the 2.0 version, it directly pulls the OpenRank rankings of all projects on GitHub, filters out projects related to large models, which not only significantly reduces the starting - point bias but also is more sensitive to the explosive power of new projects. As a result, more high - popularity and high - activity projects are discovered, and the selection threshold has also been correspondingly raised to OpenRank > 50 in the current month.

From this perspective, the 2.0 version is more in line with the original intention of Ant's "Open - Source Technology Growth" team in doing this: internally, it provides a basis for enterprise decision - making; externally, it lights up the guide for the "over - achievers" in the open - source world.

Under this new method, three major main tracks have emerged: AI Coding, Model Serving, and LLMOps.

Next, we will trace from the application layer all the way to the underlying Infra and sort out the key changes in this wave one by one.

AI Agent: AI Coding Has Gone Crazy

The AI Agent on the 2.0 panoramic map has evolved from a "treasure chest" - style tool stack to a hierarchical system similar to cloud computing —

Categories such as AI Coding, Agent Workflow Platform, Agent Framework, and Agent Tool are all available, and the professionalism and clarity have been greatly improved. The community is undergoing a process from wild growth to systematic differentiation.

The AI Agent layer in 2.0

The iteration experience of AI Agent has been like a roller - coaster ride.

Currently, the AI Agent is undergoing drastic changes on the surface, like a new continent where everyone is rushing to plant their flags. New high - popularity projects have emerged one after another in directions such as AI Coding, Chatbot, and Workflow Platform.

More interestingly, the 2.0 version has also keenly captured the signs of the deep integration of AI and the physical world — "Xiaozhi" tries to run large models on low - power chips, and Genesis provides a physical simulation platform for general - purpose robots.

Next, we will disassemble the changes in the detailed fields one by one.

1. From Crazy to Insane, the Growth Curve of AI Coding Is Still Steeply Rising

In addition to the ever - present "top - ranked" projects such as Cline, Continue, and OpenHands, new faces are constantly emerging — Gemini, marimo, Codex CLI, and OpenCode, which is positioned as a 100% open - source alternative to Claude Code. Once again, it proves that "Agent for Devs" is still the most frequent and most in - demand application scenario.

In 2025, AI Coding has completed the leap from "filling in code" to a "full - lifecycle intelligent engine": it can do more things, covering the entire chain from development to operation and maintenance; and it does things in a smarter way, supporting multi - modality, context awareness, and team collaboration.

The report predicts that the market will also release huge commercial potential accordingly — paid subscriptions, SaaS services, and value - added features will become new profit models.

AI Coding has evolved from "filling in code" to being able to do more things in a smarter way.

This trend is especially felt deeply in industry exchanges. At the Bund Summit in Shanghai this time, a guest said bluntly that there were too many AI Coding tools to use; another CEO who has been deeply involved in AI coding revealed that the AI tool reimbursement for all team members had exceeded $200.

A few months ago, AI - generated code still needed a lot of manual correction; now, the quality has soared, and only minor modifications are needed. Next, AI programming may leap from "writing code" to "dominating the entire workflow".

It is worth noting that the popularity of Gemini CLI and Codex CLI also releases the strategic signals of large companies: by binding developers through open - source toolchains, they lock them into their own closed - source model ecosystems.

This is exactly the same as Microsoft's Windows +.NET back then and Apple's iOS + Swift. Today's AI giants are using the same path to reshape a new round of developer ecosystems.

2. The Rational Return of Chatbot & Knowledge Management After the Highlight

Chatbot was the first wave of popular applications of GenAI. Cherry Studio, Open WebUI, Lobe Chat, and LibreChat reached their peaks in early 2025, receiving a lot of attention and contributions. However, the popularity did not last. After May, Chatbot entered a plateau and gradually cooled down.