-
Notifications
You must be signed in to change notification settings - Fork 2
Description
Xu hướng AI Mã nguồn mở 2026-03-16
Nguồn: GitHub Trending + GitHub Search API | Thời gian tạo: 2026-03-16 02:13 UTC
Báo cáo Xu hướng Mã nguồn mở AI — 2026-03-16
Bước 1 — Lọc
-
Các repo trending hôm nay đã chọn (liên quan AI/agent/tri thức/mô hình):
- lightpanda-io/browser (Zig) ⭐0 (+1335 today) — headless browser cho AI/automation
- volcengine/OpenViking (Python) ⭐0 (+1870 today) — context DB cho AI Agents
- shareAI-lab/learn-claude-code (TypeScript) ⭐0 (+872 today) — nano Claude Code–like agent
- obra/superpowers (Shell) ⭐0 (+1867 today) — agentic skills framework
- p-e-w/heretic (Python) ⭐0 (+1062 today) — censorship removal cho LLMs
- 666ghj/MiroFish (Python) ⭐0 (+2782 today) — swarm intelligence engine
- abhigyanpatwari/GitNexus (TypeScript) ⭐0 (+451 today) — client-side code knowledge graph + RAG
- topoteretes/cognee (Python) ⭐0 (+270 today) — knowledge engine cho agent memory
- InsForge/InsForge (TypeScript) ⭐0 (+515 today) — backend cho agentic development
- anthropics/claude-plugins-official (Python) ⭐0 (+604 today) — official Claude Code plugins
-
Từ kết quả tìm kiếm chủ đề AI (chỉ giữ dự án chuyên AI/ML): ollama/ollama, huggingface/transformers, langchain-ai/langchain, langgenius/dify, vllm-project/vllm, run-llama/llama_index, milvus-io/milvus, qdrant/qdrant, chroma-core/chroma, mem0ai/mem0, memvid/memvid, neuml/txtai, gradio-app/gradio, deepspeedai/DeepSpeed, PaddlePaddle/PaddleOCR, FlowiseAI/Flowise, etc. (Các công cụ frontend/generic bị loại trừ.)
Bước 2 — Phân loại (chọn nhóm chính)
🔧 Cơ sở hạ tầng AI
- huggingface/transformers — ⭐157,868 — Framework định nghĩa mô hình đa phương thức, chuẩn mực cho dev mô hình.
- vllm-project/vllm — ⭐73,207 — inference engine tối ưu cho throughput và bộ nhớ, quan trọng cho phục vụ LLM.
- deepspeedai/DeepSpeed — ⭐41,816 — thư viện tối ưu phân tán cho huấn luyện/inference.
- ollama/ollama — ⭐165,191 — local model runner & manager (hỗ trợ Kimi-K2.5, GLM-5, v.v.).
- open-webui/open-webui — ⭐127,333 — giao diện người dùng cho chạy mô hình/infra.
🤖 AI Agent / Workflow
- langchain-ai/langchain — ⭐129,624 — nền tảng thiết kế agent & tool-calling.
- langgenius/dify — ⭐132,919 — platform production-ready cho agentic workflows.
- shareAI-lab/learn-claude-code (trending) — ⭐28,032 (+872 today) — minimal Claude Code–style agent.
- obra/superpowers (trending) — ⭐0 (+1867 today) — agentic skills framework/methodology.
- FlowiseAI/Flowise — ⭐50,779 — xây dựng AI Agents bằng giao diện trực quan.
📦 Ứng dụng AI
- OpenBB-finance/OpenBB — ⭐63,133 — nền tảng dữ liệu và analytics có tính AI cho tài chính.
- zhagyujie/chatgpt-on-wechat — ⭐42,220 — triển khai trợ lý đa nền tảng (WeChat).
- saturndec/waoowaoo — ⭐9,392 — nền tảng sản xuất phim AI (ứng dụng ngành dọc).
🧠 LLM / Huấn luyện
- hiyouga/LlamaFactory — ⭐68,464 — fine-tuning/efficient tuning cho nhiều LLM/VLM.
- unslothai/unsloth — ⭐54,028 — công cụ fine-tuning & RL cho LLMs.
- open-compass/opencompass — ⭐6,756 — nền tảng đánh giá LLM trên bộ dataset đa dạng.
🔍 RAG / Tri thức
- run-llama/llama_index — ⭐47,690 — nền tảng agent tài liệu / RAG.
- milvus-io/milvus — ⭐43,330 — vector DB cho ANN search ở quy mô lớn.
- qdrant/qdrant — ⭐29,572 — vector DB hiệu năng cao.
- chroma-core/chroma — ⭐26,653 — search/retrieval DB mã nguồn mở.
- mem0ai/mem0 — ⭐49,936 — universal memory layer cho AI Agents.
- topoteretes/cognee (trending) — ⭐13,950 (+270 today) — knowledge engine memory.
Bước 3 — Xuất báo cáo
- Điểm nổi bật hôm nay
- Cơn sóng agent-centric tiếp tục bùng nổ: nhiều repo agentic (OpenViking, superpowers, learn-claude-code, Flowise) nổi bật với lượng sao lớn hôm nay.
- Context / memory layer là trọng tâm: OpenViking và mem0/memvid thu hút sự chú ý như giải pháp quản lý ngữ cảnh cho agent.
- Công cụ “web-automation” cho agent tăng mạnh: Lightpanda (headless browser cho AI) phản ánh nhu cầu agent thao tác web tự động.
- Hệ sinh thái Claude Code mở rộng: plugins chính thức (anthropics/claude-plugins-official) và các repo tập trung vào Claude Code nhận nhiều tương tác.
- Dự án hot theo danh mục (tiêu biểu, link + sao)
-
Cơ sở hạ tầng AI:
- huggingface/transformers — https://github.com/huggingface/transformers — ⭐157,868 — Framework model-definition chuẩn cho mô hình text/vision/audio.
- vllm-project/vllm — https://github.com/vllm-project/vllm — ⭐73,207 — Engine inference throughput/memory-efficient cho LLM serving.
- ollama/ollama — https://github.com/ollama/ollama — ⭐165,191 — Local model runner hỗ trợ nhiều model mới (Kimi-K2.5, GLM-5…).
-
AI Agent / Workflow:
- langchain-ai/langchain — https://github.com/langchain-ai/langchain — ⭐129,624 — Nền tảng agent engineering tiêu chuẩn.
- langgenius/dify — https://github.com/langgenius/dify — ⭐132,919 — Production-ready agentic workflow platform.
- shareAI-lab/learn-claude-code — https://github.com/shareAI-lab/learn-claude-code — ⭐28,032 (+872 today) — Nano Claude Code agent, lý giải spike quan tâm.
-
Ứng dụng AI:
- OpenBB-finance/OpenBB — https://github.com/OpenBB-finance/OpenBB — ⭐63,133 — Ứng dụng AI cho phân tích/đầu tư.
- zhagyujie/chatgpt-on-wechat — https://github.com/zhayujie/chatgpt-on-wechat — ⭐42,220 — Trợ lý đa nền tảng, tiêu biểu cho deploy trợ lý cá nhân.
-
LLM / Huấn luyện:
- hiyouga/LlamaFactory — https://github.com/hiyouga/LlamaFactory — ⭐68,464 — Unified efficient fine-tuning cho 100+ LLM/VLM.
- unslothai/unsloth — https://github.com/unslothai/unsloth — ⭐54,028 — Tối ưu training & RL cho LLMs.
-
RAG / Tri thức:
- run-llama/llama_index — https://github.com/run-llama/llama_index — ⭐47,690 — Leading document agent / OCR + RAG.
- milvus-io/milvus — https://github.com/milvus-io/milvus — ⭐43,330 — Vector DB cloud-native cho ANN search.
- volcengine/OpenViking — https://github.com/volcengine/OpenViking — ⭐0 (+1870 today) — Context DB cho agents, lý do spike là khả năng phân lớp context/hierarchy.
-
Phân tích tín hiệu xu hướng (200–300 từ)
Cộng đồng đang đổ vào hai lớp công nghệ đồng thời: (1) agent-centric stacks và (2) context/memory + RAG. Sự xuất hiện của nhiều repo agent (langchain, dify, Flowise, obra/superpowers, shareAI-lab/learn-claude-code) cho thấy nhu cầu xây dựng hệ thống “AI có tác vụ” (autonomous agents) tăng mạnh — không chỉ chat UI mà là pipeline tự động, orchestration và multi-tool calling. Song song đó, spike mạnh ở dự án quản lý ngữ cảnh (OpenViking, mem0, memvid, cognee, llama_index) phản ánh nhận thức rằng hiệu suất agent thực tế phụ thuộc vào lớp context: long-term memory, hierarchical context delivery và efficient retrieval. Vector DB/RAG vẫn là hạ tầng thiết yếu (Milvus, Qdrant, Chroma, LlamaIndex), nhưng xu hướng rõ rệt là tích hợp chặt vào agent workflows — memory-as-a-service thay vì pipeline thủ công. Ngoài ra, các công cụ inference/serving và local-first model runners (vllm, ollama) tiếp tục thu hút vì áp lực về chi phí, privacy và latency. Sự chú ý tới headless browser cho agent (Lightpanda) và toolkits jailbreak/defensive (heretic) còn báo hiệu hai nhu cầu song song: agent cần tương tác web an toàn và cộng đồng vẫn khám phá ranh giới an ninh/kiểm duyệt mô hình. Cuối cùng, hệ sinh thái Claude Code đang tạo một “mini-stack” plugin + best-practice, đẩy nhanh adoption cho developer tập trung vào Claude-like agents. -
Điểm nóng cộng đồng — 3–5 dự án/định hướng cần chú ý
- OpenViking (https://github.com/volcengine/OpenViking) — context DB cho agents: quan trọng để chuẩn hóa memory/hierarchy cho hệ agent.
- lightpanda-io/browser (https://github.com/lightpanda-io/browser) — headless browser thiết kế cho AI: mở ra khả năng agent thao tác web đáng tin cậy.
- mem0ai/mem0 (https://github.com/mem0ai/mem0) & memvid (https://github.com/memvid/memvid) — memory layers serverless cho agents: giảm phức tạp RAG.
- langchain-ai/langchain (https://github.com/langchain-ai/langchain) — agent engineering: vẫn là tiêu chuẩn để xây workflow agent.
- vllm (https://github.com/vllm-project/vllm) — inference engine: tối ưu hóa phục vụ LLM với chi phí/latency thấp.
Kết luận ngắn: Hôm nay thể hiện rõ thế hệ “agent + memory + RAG + efficient inference” — xây dựng agent thực tế đang là trọng tâm của cộng đồng mã nguồn mở AI.
Bản tin này được tạo tự động bởi agents-radar.