I went down a rabbit hole this morning scrolling through the 2025 and early-2026 Juejin roundups — the front-end dev tool lists, the monthly GitHub trending recaps, the open-source AI ecosystem overviews — and the thing that finally crystallized for me is that the open-source AI stack is starting to look like a real full stack, not a collection of cool demos. That is a different sentence than I would have written six months ago, and I want to put it somewhere I can find it.
The piece that pushed me over the edge was the 2025-2026 AI open-source panorama post. Reading it back to back with the monthly GitHub trending lists from May, June, and October 2025, the same names keep showing up in the same rough layers. At the model layer you have LLaMA 3, the Phi series from Microsoft, LongCat-Flash-Chat from Meituan doing the MoE thing, and OpenAI's GPT-OSS-120B and 20B which were the first open weights from them since GPT-2 and which I genuinely did not see coming. At the agent framework layer it is CrewAI, OpenDevin, LangGraph, AutoGen Studio, and Semantic Kernel fighting for the same mindshare. At the memory and RAG layer it is mem0, DB-GPT, and the LangChain ecosystem generally. At the multimodal and creative layer it is ComfyUI, Stable Diffusion Web UI, and the DALL-E 3 reproduction projects. At the deployment and orchestration layer it is n8n, Hugging Face Transformers, and the ONNX Runtime family. The same names, the same rough layers, every month, and I am a little skeptical of any list that clean because curation bias is real, but the overlap is doing work.
The thing I keep coming back to is how different this looks from a year ago. A year ago the open-source story was a few labs publishing weights and a thriving demo scene on top of them. Now there is a recognizable full stack forming, and the October trending list is a good snapshot of where the pressure points are. Agent-S from Simular AI is pushing on the computer-use wall, mem0 and supermemory are pushing on the persistent memory wall, winboat is pushing on the cross-platform desktop wall, the Claude cookbooks and prompt engineering tutorial from Anthropic are pushing on the developer onboarding wall, and nanoGPT is still the cleanest way for me to explain to a junior what a transformer is actually doing. To be fair, I have not stress-tested most of these in production, and I want to actually try Agent-S on my own machine before I oversell it, but the OSWorld numbers from late 2025 are the first time I have seen computer-use feel like more than a demo reel.
The meta-pattern that jumped out of the comparison between the May and October trending lists is that the open-source ecosystem is splitting into two camps in a way that mirrors the closed-source world. Camp one is the polished end-user product you actually open every day — Cursor, Claude Code, GitHub Copilot, 通义灵码, Antigravity. Camp two is the infrastructure layer that the polished products are quietly built on top of, and the projects landing in the trending page are mostly camp two. The closed-source tools get the Juejin headlines, the open-source infrastructure gets the GitHub stars, and the people shipping production agents are reading both.
What I think this means practically, at least for me, is that the next time I am picking a tool, I am going to spend more time reading the open-source infrastructure layer than I used to. The polished editor layer has converged enough that the differences between Cursor and Claude Code on any given Tuesday are smaller than the differences between having or not having a working memory layer underneath. ComfyUI taught me that a year before it became obvious, and I expect Agent-S and mem0 to do the same thing in 2026. Honestly, I think the open-source community is closer to owning the infrastructure layer than the application layer right now, and that is fine, but it is also the part most engineers I talk to are still underweighting.
I will reassess in three months again, because that is the only honest timeline I can commit to, and the last time I said that I was mostly bouncing between Cursor and Claude Code which is still where I land. What has changed is that I now treat the open-source infrastructure layer as something I read on purpose, not something I stumble into. The 2025 roundups made the stack visible, and I think the 2026 roundups are going to make it feel normal. That feels like the right direction.
Top comments (0)