Why TensorFlow Is Dying and What the New AI Open‑Source Landscape Looks Like

Only 100 Days, the Open‑Source World Changes Dramatically

TensorFlow, the once‑dominant open‑source framework, is now on the brink of extinction. Data shows that after a ten‑year peak, its community activity has fallen irreversibly to a level lower than at its launch.

In stark contrast, PyTorch has been climbing a steep, red‑hot curve.

This observation comes from Wang Xu, Vice‑Chair of Ant Group’s Open‑Source Technology Committee, who presented a sobering trend analysis at the recent Bund Conference.

In Ant’s latest “Large Model Open‑Source Development Ecosystem Panorama 2.0”, TensorFlow has been officially removed.

The conclusion is not baseless; it is derived from a calm data‑driven insight.

Over a decade, TensorFlow’s community activity peaked but then declined irreversibly to its lowest point, even lower than at its inception.

Conversely, PyTorch’s activity has surged.

Wang Xu explained that the panorama is built on a data‑driven project called OpenRank , an influence‑assessment algorithm similar to PageRank but applied to open‑source communities. It evaluates a project’s influence by analyzing the collaboration network among projects and developers.

In version 2.0, Ant upgraded the methodology: they directly pull the monthly OpenRank rankings of all GitHub projects and set an admission threshold of “OpenRank > 50” to filter projects belonging to the large‑model ecosystem.

This adjustment led to noticeable changes: 39 new projects were added (a 35% replacement rate) and 60 existing projects were removed.

Wang Xu believes such rapid turnover is expected in the fast‑iteration AI field.

Which projects get removed? Primarily short‑term hype projects, such as the open‑source replicas OpenManus and OWL that rose with the Manus hype in March but lost community activity after the hype faded.

Second, projects with slow iteration speed, e.g., NextChat, which lagged behind newer clients like Cherry Studio and LobeChat, leading to user and developer loss.

Third, laggards in niche competition. In the realm of edge‑model deployment, MLC‑LLM and GPT4All were overtaken by Ollama, which offers a more complete ecosystem and user experience.

These observations illustrate the fierce competition in the AI open‑source ecosystem.

New Paradigm: Open‑Source Redefined

The definition of open‑source is becoming more complex in the AI era. Examining the top‑10 most active projects in the 2.0 panorama reveals that some high‑activity projects do not use OSI‑approved licenses:

Dify (rank 4) – adds multi‑tenant usage restrictions and brand‑logo protection on top of Apache 2.0.

Cherry Studio (rank 7) – adopts a dual‑license model based on organization size, requiring commercial licensing for larger teams.

n8n (rank 9) – uses its own “Sustainable Use License” that limits commercial distribution.

Although these modifications deviate from strict license definitions, they may help achieve a more balanced benefit distribution and ensure the long‑term health of the ecosystem.

The shift reflects two paradigm changes:

1. Open‑source’s operational attribute is enhanced; GitHub becomes a crucial go‑to‑market (GTM) channel. GitHub now serves not only as code hosting but also as a platform for product releases, user feedback, and community marketing. Even closed‑source commercial products (e.g., Cursor, Claude‑Code) maintain active GitHub communities for interaction and promotion.

Consequently, the degree of code openness matters less, while community activity (stars, issues, PRs) becomes a key indicator of product vitality and market acceptance.

2. A new balance between “community openness” and commercial interests. New AI open‑source projects define commercial goals early and use customized license terms to capture community benefits while protecting core business interests. For example, Dify’s multi‑tenant restriction guides SaaS‑style services, and n8n’s license steers users toward commercial pathways.

New Battlefield: From Framework Wars to Performance‑King

In the 1.0 stage, competition centered on broad functionality exploration. In the 2.0 stage, the focus has shifted clearly.

Technical field development trends show a decline in Agent Frameworks (e.g., LangChain, LlamaIndex, AutoGen) while Model Serving and AI Coding exhibit significant growth.

This reflects the AI industry’s transition from an exploratory phase to an engineering‑deployment phase.

Agent frameworks’ activity drop does not mean the concept failed; rather, the market moves from broad exploration to rational selection, emphasizing performance, stability, and cost.

Thus, the focus has moved from “can it be done?” to “can it run efficiently, economically, and reliably?”

High‑performance inference engines such as vLLM, SGLang, and NVIDIA’s TensorRT‑LLM dramatically improve GPU utilization and throughput, becoming the core of AI infrastructure and directly influencing the commercial feasibility of upper‑layer applications.

New Pattern: The US and China Lead the Open‑Source World

Globally, developers from the United States (24% of identifiable contributors) and China (18%) dominate. Based on OpenRank contribution scores, the US accounts for 37.4% and China 18.7%, together exceeding 55% of total contribution.

In the AI Infra domain, the US leads with 43.39% contribution versus China’s 22.03%. In the AI Agent domain, the gap narrows: the US contributes 24.62% and China 21.5%.

This “dual‑center” pattern shows the US leading in foundational infrastructure, while China shows strong momentum in application‑level innovation.

One More Thing

Contrasting TensorFlow’s decline, new open‑source projects rise rapidly. For example, the AI Coding project OpenCode (an open‑source alternative to Claude Code) and Google’s Gemini CLI gained massive community attention within months.

The most noteworthy newcomer is Browser‑use – built by two graduate students in nine months and already earning 60 K stars.

So, in another 100 days, which project will become the next “Browser‑use” sensation?