The 100-day suspense has been resolved. On the morning of September 13, Ant Group’s open-source team (“Open Source Technology Growth”) presented version 2.0 of its LLM Open-Source Development Ecosystem Landscape at the Bund Summit in Shanghai. Three months ago the project called itself “a real-world hackathon, broadcast live” — and the label still fits. The earlier edition captured the ecosystem’s infancy; this one shows a different face entirely.
Browse it at antoss-landscape.my.canva.site. This edition covers 114 projects across 22 fields — 21 fewer than before. Thirty-nine are newcomers, while 60 have left the stage, including one-time star royalty: NextChat, OpenManus, FastGPT, and GPT4All, overtaken after slow iteration and fading communities. The most dramatic exit is TensorFlow’s final bow — the former superstar could not withstand PyTorch’s assault, and PyTorch now rules alone.
The overall trend is unmistakable: the ecosystem is undergoing a violent reshuffle. Like the Cambrian explosion, the agent layer churns hardest, with new species appearing constantly. The metabolism shows in the numbers — counting the eliminated, the LLM ecosystem’s median project age is just 30 months, average lifespan under three years, an extremely young jungle. Since the “GPT moment” (October 2022), 62 percent of today’s projects were born — twelve of them as recently as 2025. New arrivals and departures land roughly every quarter. And these young projects command unprecedented attention: average stars near 30,000, far beyond what same-aged projects ever earned before, with the top ten covering nearly the full model-ecosystem chain. The keyword cloud tells the same story: AI, LLM, Agent, and Data are the biggest, brightest letters.
The taxonomy changed too. Version 1.0’s coarse Infrastructure/Application split has evolved into three boards — AI Agent, AI Infra, AI Data — sketching the industry’s center of gravity (agents) and its direction of travel. If 1.0 still smelled of traditional open-source software ecology, 2.0 has the temperament of the agent era. From a global community of 366,521 developers, China and the US together contribute over 55 percent, with America first at 37.41 percent. The US leads AI Infra decisively — 43.39 percent, more than double China’s — and dominates AI Data even more clearly. In the application layer (AI Agent), the two run nearly level at 21.5 and 24.62 percent, reflecting Chinese developers’ heavier agent investment.
The Methodology Evolved
Why lead with methodology? Because which projects made the 2.0 map depended on it. Version 1.0 started from known head projects — PyTorch, vLLM, LangChain — and extended outward through their dependencies; the starting points set the boundaries, with an entry bar of monthly OpenRank (the X-lab influence metric) at 10 or more. Version 2.0 instead pulls OpenRank rankings across GitHub’s full universe, filters for LLM-related projects, slashes starting-point bias, and catches new-project explosions far more sensitively — raising the entry bar to monthly OpenRank above 50. Under the new method, three main tracks emerged: AI Coding, Model Serving, and LLMOps. The survey then works from the application layer down to infrastructure.
AI Agent: AI Coding Has Gone Feral
The agent layer on the 2.0 map has evolved from a toolbox pile into something resembling cloud computing’s layered structure — AI Coding, Agent Workflow Platforms, Agent Frameworks, and Agent Tools now categories in their own right, as the community moves from wild growth to systematic differentiation. The iteration has been a roller coaster: the ground shifts constantly, everyone is planting flags, and new high-heat projects keep surfacing across coding, chatbots, and workflow platforms. The 2.0 edition even caught AI reaching into the physical world — “Xiao Zhi” running large models on low-power chips, and Genesis building a physics simulation platform for general robotics.
AI Coding: from mad to possessed, the growth curve still steepening. Beyond the perennial chart-toppers Cline, Continue, and OpenHands, newcomers keep arriving — Gemini, marimo, Codex CLI, and OpenCode, positioned as a 100 percent open-source Claude Code replacement. The evidence reconfirms that “agents for developers” remains the highest-frequency, hardest-need application. 2025’s AI coding has crossed from “completing code” to a full-lifecycle intelligence engine: more of the chain covered, from development through ops, and smarter methods — multimodal, context-aware, team-coordinated. The report expects the commercial potential to follow: paid subscriptions, SaaS services, and value-added features as the new revenue models. The ground truth from industry conversations matches: at the Bund Summit one speaker admitted the coding tools have multiplied beyond anyone’s ability to try them all; another CEO disclosed his team’s AI-tool reimbursements now exceed $200 per person. A few months ago AI code needed heavy human correction; now quality has jumped to where light edits suffice. The next step may be AI coding leaping from “writing code” to “driving the whole workflow.” Note the strategic signal in Gemini CLI’s and Codex CLI’s popularity: the giants are using open toolchains to bind developers into their closed-model ecosystems — Windows-plus-.NET and iOS-plus-Swift replayed for the AI age.
Chatbots and knowledge management: rational return after the spotlight. Chatbots were GenAI’s first hit category — Cherry Studio, Open WebUI, Lobe Chat, and LibreChat peaked in early 2025 with enormous attention and contributions. But the heat did not hold; after May the category plateaued. Lobe-Chat’s daily-release cadence turned the ChatGPT conversation into something anyone could self-host, yet as features matured, users confronted the truth that a chatbot’s core value is conversation — and conversation has limited productivity upside. Cherry Studio opened another path, fusing chatbot with knowledge bases and notes as a “personal knowledge assistant,” closer to long-term productivity needs — likely why it made the list against the trend, the second-most-active newcomer at seventh overall. The cooling is not decline: Lobe-Chat still holds 64.7k stars with daily releases. Chatbots simply stopped being the only show.
Agent workflow platforms consolidating into real infrastructure. When GPT-5, Claude, and Gemini converge on capability, differentiation migrates to memory — which spawned a wave of RAG-plus-agent projects. RAGFlow, MaxKB (knowledge-base Q&A focus), FastGPT (lightweight RAG), and Flowise AI (visual RAG-based QA workflows) all rode the wave; though the trend has cooled, July’s OpenRank of 1,418 remains far above January’s start. Dify’s positioning is the most complete: RAG plus agent workflows plus application publishing, user management, and observability — a prototype-to-production platform clearly leading in community influence and commercialization. Meanwhile the frontier is pushing past classical RAG: Letta (not yet listed) experiments with offline learning, letting agents genuinely learn from experience rather than merely retrieve — hinting that agent workflows may evolve from retrieval augmentation toward long-term learning.
Agent tools: hit after hit. Agent tools patch the model’s hard deficiencies, moving AI from “can chat” to “can act” — the hottest innovation track of the moment. LiteLLM, Supabase, Vercel, ComfyUI, and mem0 each own a critical link: inference routing, data storage, memory management, external interaction. mem0 gives agents long-term memory with context awareness; Supabase is evolving into the GenAI data infrastructure, solving real-time information and memory gaps; and Browser-use took 60,000 stars in nine months by teaching agents to operate web pages — the landmark project of automation’s last mile. Agent frameworks, meanwhile, fell hard: four of the steepest decliners are orchestration frameworks — Eliza, LangChain, LlamaIndex, AutoGen — as head projects trim community investment and attention migrates from general frameworks to application-specific innovation in memory, tool calling, and interfaces. But Camel-AI, CrewAI, Agno, and ELIZA.OS rising says collaborative multi-agent systems are drawing more attention, and LiveKit Agents’ entry marks real-time interaction — voice, multi-party — a zone the map barely covered before.
AI Infra: Serving Still Burning, LLMOps Inherits MLOps
Less explosive than the agent layer, the infra layer still shows strong undercurrents: relentless cloud-inference competition, blooming lightweight inference, and increasingly specialized LLMOps. Model serving stays the main battlefield — the only track that can stand toe to toe with AI Coding. Deployment lives or dies on inference efficiency, cost, and safety; the track has run hot since 2023. vLLM and SGLang lead high-performance cloud inference with expanding ecosystems, vLLM’s stability keeping its heat firm. The giants are piling in — NVIDIA’s TensorRT-LLM stays hot and the newer Dynamo pairs with it, deep-binding GPU hardware and treating the inference framework as a new handle for compute dominance. Local inference is sprinting too: ollama flew large models from expensive clouds into ordinary homes, running on personal computers and even mobile devices; its heat has dipped slightly of late, but GPUStack and ramalama are quickly filling the lane — the “lightweight plus cloud-optimized” path keeps minting players.
LLMOps inherits MLOps as 2025’s fastest-growing main track: Phoenix, Langfuse, Opik, Promptfoo, 1Panel, and Dagger cover nearly everything from monitoring and prompt evaluation to workflow management. The key shift is definitional — work once split between model evaluation and traditional MLOps is being absorbed into whole-life-cycle LLM management, and the focus moves from the small-model era’s training accuracy and data pipelines to “how to keep models running stably and controllably.” LLMOps now centers on monitoring, prompt effectiveness, observability, and safety — the support structure that carries large models into real applications. Model training, by contrast, fell below its January level. The retreat is rational: different industries want different model flavors, training a GPT-4-class model from scratch costs too much, and the community pivoted to low-cost fine-tuning — Unsloth integrating QLoRA to run on mid-range hardware. The application layer became the new battlefield, friendlier to developers with more room to contribute; PyTorch-era training frameworks are mature and plateaued, and while Swift, Unsloth, and LLaMA-Factory drew short-term attention, none has broken through sustainably. Meanwhile Ray dominates distributed computing, and the rise of specialist kernel libraries — TransformerEngine, DeepEP, Triton, Modular — marks large models spawning their own software stack, from Python APIs down to operators, purpose-built for LLMs.
AI Data: Calm Before the Larger Storm
Compared with the churn above it, AI Data is remarkably calm. The chain is complete — annotation, integration, governance, vector databases and search, up through application frameworks — covering essentially every data-processing stage. Most projects are old faces born between 2014 and 2020, built for the AI 1.0 (classical ML) era; Chroma is among the few newcomers to catch the LLM wave, arriving in 2022 exactly at the RAG breakout. But the heat keeps draining. Part of that is maturity — Iceberg and Delta Lake data lakes, Milvus, Weaviate, and Elasticsearch vector stores are proven infrastructure puzzles, with no obvious new breakthroughs. Part is enterprise orientation — individual developers rarely feel the value directly, so GitHub attention cannot sustain. But calm often precedes the bigger storm. AI Data’s real test is emerging in the natively-multimodal era: vectorize everything, RAG at scale, multimodal data processing — experts predict data platforms will evolve from “warehouses” into “hubs”: intelligent, dynamic, real-time systems continuously feeding real-world perception to the model brain. The old frameworks remain solid but are being pushed toward their limits; the next generation of LLM-native data infrastructure is quietly gestating.
The Business Game: ‘Open Source’ Turning Strange
Before the 2.0 map shipped, a subtle trend was visible: more and more LLM projects abandon MIT and Apache 2.0 for self-drafted “open-source license agreements,” carving gray zones between openness and control. The new-style licenses reserve more intervention rights for the grantor — where traditional licenses are irrevocable once granted, some LLM licenses remove that constraint, leaving room for later commercial games. Others cap by user scale: exceed a monthly-active-user threshold and licensing becomes a negotiation. The 2.0 data confirms the observation — most newcomer licenses are restrictive variants, igniting “is this even open source” debates. Dify uses a BSL (Business Source License) variant — code opens now, flips permissive years later, protecting commerce meanwhile. n8n uses fair-code, explicitly framed as “stopping the giants from freeloading.” Cherry Studio’s terms require separate authorization for commercial use. And a second blur comes from GitHub itself: Cursor and Claude Code keep even their code closed while trending hot on the platform as feedback-collection windows — what counts as an “open-source community” is itself being challenged by this wave.
Bonus: The Model Battlefield
As a bonus, the 2.0 edition charted every major model release — open and closed, domestic and foreign — since January 2025, annotated with parameters and modalities. Five patterns emerge. Route divergence, open versus closed: Chinese open models keep blooming while top foreign labs keep betting closed; Meta, which once counter-punched with Llama, is tightening its open cadence — Zuckerberg’s open letter admits “more careful choices about what to open” — and Llama 4’s reception made the position awkward. It echoes Alibaba Cloud founder Wang Jian’s Bund Summit judgment: the open-versus-closed choice has become a key variable of AI competition. MoE everywhere, parameters sprinting to trillions: DeepSeek, Qwen, and Kimi flagships all ship mixture-of-experts; sparse activation lets total parameters soar to trillion scale while activating only a slice at inference — K2, Claude Opus, and o3 arrived in sequence, performance jumping while training and inference compute hit new highs. Reasoning is the new standard feature: last year’s contest was scale; this year’s is reasoning. DeepSeek R1’s reinforcement-learning gains made automated reasoning and complex decisions a launch requirement, and Qwen, Claude, and Gemini now ship fast-think/slow-think hybrid modes that switch response speeds between tasks like people do. Multimodality erupts: the strongest brains stopped staring at language — language, image, and voice interaction are now table stakes, with Chinese vendors leading on both open and closed lines, and the voice ecosystem (Pipecat, LiveKit Agents, CosyVoice) rapidly completing its toolchain. Distance to mature video modality — let alone AGI — remains real. Evaluation pluralizes: some still trust human-voting arenas like Design Arena and LMArena; others trust objective test suites with standard answers. Meanwhile API-gateway platforms like OpenRouter, sitting on natural call-volume data, are forming “data-voting” leaderboards — and LiteLLM’s surging activity suggests call statistics may become a new evaluation instrument in their own right. Source: Synced (machine.ai), edited by Wu Xin; the Ant Open-Source Ecosystem Landscape 2.0.
