69个最佳开源AI仓库推荐（2026年4月）

✍ self.dll🕐 2026-04-29📦 24.7 KB 🟢 已读 𝕏 文章列表

文章精选了69个经过实战验证、维护活跃的开源AI仓库，旨在替代昂贵的API订阅。内容涵盖完整的AI技术栈，包括本地大模型推理、RAG知识库构建、AI智能体框架、提示词与评估、模型微调、工具链及部署等十大领域，为开发者、研究人员和独立黑客提供免费的生产级AI工具清单。

开源 AI工具 LLM Agent RAG 模型微调 Ollama LangChain 本地部署资源合集

# 69 Best Open-Source AI Repositories in April 2026 **作者**: self.dll **日期**: 2026-04-28T19:47:49.000Z **来源**: [https://x.com/seelffff/status/2049214021430325677](https://x.com/seelffff/status/2049214021430325677) --- ![Image](/doc/1264/image/img_53da7889e793.jpg) the average AI startup pays $8,000–$50,000/year in tool subscriptions. the people actually building AI - researchers, engineers, indie hackers - use open-source. they always have. i went through 300+ repos. pulled out 69 that are production-ready, actively maintained, and genuinely useful. this is the full AI stack. free. ## This is the toolbox. what's inside: - 01 llm inference - run models locally. no api. no limits. no bill. - 02 rag & knowledge - make ai answer from your own data. - 03 ai agents - let the model act, not just answer. - 04 prompts & evals - stop guessing. start measuring. - 05 fine-tuning -make the model yours. - 06 tools & context -feed the model what it actually needs. - 07 deployment -ship it. then scale it. - 08 claude-specific - if you use claude, these are not optional. - 09 data prep - garbage in, garbage out. fix the input. - 10 vision & multimodal - beyond text. - 69 bonus - don't want to run all this yourself? ![Image](/doc/1264/image/img_9fd583c063b5.jpg) # 01 llm inference run models locally. no api. no limits. no bill. ## 01 ollama ★ 98K - github.com/ollama/ollama run llama, mistral, qwen, gemma locally with one command. the fastest way to get a model running on your machine. `ollama run llama3` and you're done. supports gpu acceleration, rest api, openai-compatible endpoints. ## 02 llama.cpp ★ 72K - github.com/ggml-org/llama.cpp llm inference in pure c++. runs on cpu, gpu, apple silicon. the engine behind most local ai tools. if ollama is the car, llama.cpp is the engine. extremely fast, low memory usage. ## 03 vllm ★ 44K - github.com/vllm-project/vllm high-throughput llm serving engine for production. continuous batching, paged attention, openai-compatible api. the standard for deploying models at scale. used by most serious ai companies. ## 04 lm studio ★ 28K desktop app for running local llms with a clean ui. download models from hugging face, run them locally, get an openai-compatible local server. best onboarding for non-developers. github.com/lmstudio-ai/lmstudio.js ## 05 jan ★ 26K - github.com/janhq/jan open-source chatgpt alternative that runs 100% offline. clean ui, model management, local api server. works on mac, windows, linux. no data leaves your machine. ## 06 text-generation-webui ★ 42K - github.com/oobabooga/text-generation-webui the swiss army knife for local llms. supports every model format, every backend, every sampler. character mode, notebook mode, api mode. the most feature-complete local ui that exists. ## 07 localai ★ 26K - github.com/mudler/LocalAI self-hosted openai drop-in replacement. same api, local models. swap claude/gpt with local llms in any app without changing a single line of code. # 02 rag & knowledge make ai answer from your own data. ## 08 langchain ★ 98K - github.com/langchain-ai/langchain the most popular llm framework. chains, agents, retrievers, memory. connects llms to any data source or tool. massive ecosystem of integrations. if you're building anything with ai, you'll hit langchain eventually. ## 09 llamaindex ★ 38K - github.com/langchain-ai/langchain data framework for llm applications. index any data source - pdf, sql, notion, slack - and query it with natural language. better than langchain for pure rag use cases. ## 10 rag-anything ★ 12K - github.com/HKUDS/RAG-Anything multimodal rag for claude and other llms. handles text, tables, images, charts, graphs. not just pdfs - actually everything. 6 lines to set up. used in production by serious teams. ## 11 chroma ★ 16K - github.com/chroma-core/chroma the open-source vector database. store embeddings, search by similarity, filter by metadata. runs in-memory or persistent. the simplest way to add semantic search to any project. ## 12 weaviate ★ 12K - github.com/weaviate/weaviate vector database with built-in ml models. hybrid search, multi-tenancy, real-time updates. production-ready, scales to billions of objects. used by companies you've heard of. ## 13 haystack ★ 18K - github.com/deepset-ai/haystack end-to-end nlp framework for rag pipelines. modular, production-ready, works with any llm or vector db. the most mature rag framework available. ## 14 docling ★ 22K - github.com/DS4SD/docling convert documents to structured markdown for ai. handles pdfs with tables, figures, formulas - not just plain text extraction. built by ibm research. # 03 ai agents let the model act, not just answer. ## 15 autogen ★ 40K - github.com/microsoft/autogen multi-agent conversation framework by microsoft. agents talk to each other, delegate tasks, write and execute code. the most powerful framework for complex agentic workflows. ## 16 crewai ★ 28K - github.com/crewAIInc/crewAI orchestrate role-playing ai agents. define a crew, assign roles, set goals — agents collaborate like a team. easiest way to build multi-agent systems that actually work. ## 17 langgraph ★ 10K - github.com/langchain-ai/langgraph build stateful multi-agent workflows as graphs. nodes are agents or functions, edges are transitions. handles complex logic, loops, human-in-the-loop. the production-grade agent framework. ## 18 agno ★ 22K - github.com/agno-agi/agno build fast multi-modal ai agents. supports any llm, any tool, memory, knowledge, storage. 10x faster than langchain for simple agents. clean api, excellent documentation. ## 19 smolagents ★ 14K - github.com/huggingface/smolagents minimal agent framework by hugging face. code agents that write and execute python to solve tasks. incredibly simple - 1000 lines of code total. the anti-langchain. ## 20 openhands ★ 48K - github.com/All-Hands-AI/OpenHands open-source devin alternative. ai software engineer that writes code, runs tests, fixes bugs, deploys. works with claude, gpt-4, local models. the most capable coding agent. ## 21 superagi ★ 16K - github.com/TransformerOptimus/SuperAGI self-hosted autonomous ai agent infrastructure. agent marketplace, performance telemetry, concurrent agents, graphical ui. run multiple agents in parallel on your own server. ## 04 prompts & evals stop guessing. start measuring. ## 29 dspy ★ 22K - github.com/stanfordnlp/dspy programming - not prompting - llms. define what you want, dspy optimizes the prompts automatically. from stanford nlp. replaces manual prompt engineering with systematic optimization. ## 30 guidance ★ 20K - github.com/guidance-ai/guidance control llm output structure with code. interleave generation with logic, force json schemas, constrain outputs. when you need the model to output exactly what you need. ## 31 outlines ★ 11K - github.com/dottxt-ai/outlines structured text generation. force models to output valid json, regex patterns, specific formats. zero prompt engineering needed - guaranteed output structure. 32 promptfoo ★ 6K - github.com/promptfoo/promptfoo test and eval your prompts. run automated tests, compare models, catch regressions. like unit tests but for ai. essential before shipping anything to production. ## 33 braintrust ★ 3K - github.com/brainlid/langchain eval framework for llm apps. track quality across model versions, prompts, and configurations. because vibes aren't a metric. ## 34 instructor ★ 9K - github.com/instructor-ai/instructor structured outputs from llms using pydantic. define a schema, get back a validated python object. works with openai, anthropic, google, local models. the cleanest structured output solution. # 05 fine-tuning make the model yours. ## 35 unsloth ★ 24K - github.com/unslothai/unsloth fine-tune llms 2x faster, 80% less memory. supports llama, mistral, qwen, gemma. runs on a single gpu. the only fine-tuning library you need if you're resource-constrained. ## 36 axolotl ★ 8K - github.com/axolotl-org/axolotl streamlined fine-tuning for llms. yaml config, every dataset format, every training technique. the ops layer on top of hugging face transformers. used by most serious fine-tuners. ## 37 llama-factory ★ 40K - github.com/hiyouga/LLaMA-Factory fine-tune 100+ llms with zero code. web ui, supports lora, qlora, full fine-tuning. the most user-friendly fine-tuning tool available. 40K stars for a reason. ## 38 trl ★ 12K - github.com/huggingface/trl transformer reinforcement learning. rlhf, dpo, ppo - the techniques used to align gpt-4 and claude. by hugging face. for when you want to train models to do what you actually want. ## 39 torchtune ★ 5K - github.com/pytorch/torchtune pytorch-native fine-tuning library from meta. simple, hackable, well-documented. the reference implementation for fine-tuning in pure pytorch. ## 40 mergekit ★ 4K - github.com/arcee-ai/mergekit merge multiple fine-tuned models into one. slerp, ties, dare, linear merge - all the techniques. no gpu needed for merging. create frankenstein models that outperform their parents. used by everyone releasing merged models on hugging face. # 06 tools & context feed the model what it actually needs ## 41 markitdown ★ 38K - github.com/microsoft/markitdown convert any file to markdown. pdf, word, excel, powerpoint, images, audio. feeds clean structured text to your llm instead of garbage. by microsoft. 38K stars. ## 42 files-to-prompt ★ 3K - github.com/simonw/files-to-prompt turn your entire codebase into one prompt. respects .gitignore, recursive, filterable. by simon willison. the simplest tool for feeding projects to claude. ## 43 crawl4ai ★ 30K - github.com/unclecode/crawl4ai web scraping for ai. extracts clean markdown from any url, handles js-heavy sites, structured data extraction. the web data layer for any ai pipeline. ## 44 firecrawl ★ 25K - github.com/mendableai/firecrawl turn any website into llm-ready data. full site crawling, structured extraction, clean markdown output. the production-grade web scraper for ai apps. ## 45 playwright-mcp ★ 31K - github.com/microsoft/playwright-mcp give claude a real browser. navigate, click, screenshot, read dynamic content. analyze any site in 30 seconds. the most powerful mcp server for web tasks. ## 46 model-context-protocol ★ 11K - github.com/anthropics/model-context-protocol the standard for connecting claude to external tools. official anthropic mcp. plug in any api, database, service. hundreds of servers in the ecosystem. ## 47 awesome-mcp-servers ★ 27K - github.com/punkpeye/awesome-mcp-servers 500+ ready-made mcp servers. github, slack, notion, databases, browsers, finance. every integration you'll ever need in one catalog. ## 48 n8n ★ 47K - github.com/n8n-io/n8n self-hosted workflow automation with 400+ integrations. connect llms to any app. trigger ai workflows on schedules or webhooks. run custom js/python in nodes. the automation layer behind serious ai pipelines. replaces $50K/year zapier. # 07 deployment ship it. then scale it. ## 49 litellm ★ 16K - github.com/BerriAI/litellm one api for 100+ llms. openai format, works with claude, gpt, gemini, local models. load balancing, fallbacks, cost tracking. the proxy layer between your app and every llm provider. ## 50 bentoml ★ 7K - github.com/bentoml/BentoML build and deploy ai services. package models, create apis, deploy anywhere. from local testing to production kubernetes. the mlops layer that doesn't require a devops team. ## 51 ray serve ★ 34K - github.com/ray-project/ray distributed ai inference at scale. serve multiple models, autoscale, handle millions of requests. used by openai, anyscale, production ai companies. overkill until you need it. ## 52 triton inference server ★ 8K - github.com/triton-inference-server/server nvidia's production inference server. maximum gpu utilization, dynamic batching, multi-model serving. the standard for gpu inference in enterprise. ## 53 lorax ★ 3K - github.com/predibase/lorax serve hundreds of lora fine-tuned models on one gpu. one base model, hundreds of adapters loaded dynamically. 10x cost reduction for serving fine-tuned models. ## 54 supabase ★ 73K - github.com/supabase/supabase the default backend for ai applications. open-source firebase alternative built on postgres. real-time database, auth, storage, edge functions, vector search. 73K stars. replaces firebase + auth0 ($15K/year). # 08 claude-specific if you use claude, these are not optional. ## 55 obra/superpowers ★ 160K - github.com/obra/superpowers adds superpowers to claude code. deep code analysis, auto-refactor, project-wide editing. works as a layer on top of the official cli. 160K stars. the most popular claude enhancement. ## 56 claude-code-skills ★ official - github.com/anthropics/claude-code-skills official anthropic skills framework. skill.md patterns that teach claude to handle documents, automations, workflows without errors. the foundation of how claude code handles complex tasks. ## 57 free-claude-code ★ 2K - github.com/Alishahryar1/free-claude-code run claude code completely free via github models api. trending #1 on github. step by step guide + setup scripts. $0. forever. ## 58 claude-mem ★ 1K - github.com/thedotmack/claude-mem persistent memory for claude. auto-captures everything claude does across sessions. replaces paid context management tools. claude remembers who you are and what you're working on. # 09 data prep garbage in, garbage out. fix the input. ## 59 unstructured ★ 10K - github.com/Unstructured-IO/unstructured extract and transform unstructured data for llms. pdfs, html, word, images, emails - all parsed into clean chunks ready for rag. the data layer most ai pipelines are missing. ## 60 datatrove ★ 3K - github.com/huggingface/datatrove large-scale data processing for llm training. by hugging face. process terabytes of text with deduplication, quality filtering, and content classification. what the big labs use. ## 61 trafilatura ★ 3K -github.com/adbar/trafilatura web content extraction for ai. strips boilerplate, keeps content, outputs clean text or markdown. the best single-page web extractor for feeding text to models. 62 semchunk ★ 1K - github.com/umarbutler/semchunk semantic text chunking for rag. splits text at natural boundaries instead of arbitrary token counts. better chunks → better retrieval → better answers. ## 63 datachain ★ 2K - github.com/iterative/datachain ai-native dataset management. version, query, and transform multimodal datasets. works with images, video, text, embeddings. built for llm training workflows # 10 vision & multimodal beyond text. ## 64 moondream ★ 10K - github.com/vikhyat/moondream tiny vision language model that runs anywhere. 1.6B parameters. describe images, answer visual questions, detect objects. runs on a raspberry pi. the smallest useful vision model. ## 65 internvl ★ 7K - github.com/OpenGVLab/InternVL state of the art open-source vision model. matches gpt-4v on most benchmarks. understand images, charts, documents, screenshots. the open alternative to claude's vision. ## 66 whisper ★ 74K - github.com/openai/whisper open-source speech recognition by openai. transcribes audio in 99 languages. runs locally, handles accents, background noise, technical jargon. feed audio to your llm pipeline. ## 67 insanely-fast-whisper ★ 8K - github.com/Vaibhavs10/insanely-fast-whisper whisper but 10-20x faster. one command, automatic gpu optimization, batch processing. transcribe a 2-hour podcast in 2 minutes on consumer hardware. ## 68 stable-diffusion-webui ★ 143K - github.com/AUTOMATIC1111/stable-diffusion-webui the browser interface for stable diffusion. generate, edit, upscale images from text. 143K stars - the most starred ai repo outside llms. hundreds of extensions, styles, controlnet, inpainting. runs on your gpu. # 69 kreo - copy the top polymarket traders ★ - http://t.me/KreoPolyBot?start=ref-kreohub kreo tracks the top-performing wallets on polymarket in real time and automatically copies their trades. no code, no monitoring, runs 24/7. the top polymarket traders use sophisticated models to find mispriced markets -kreo lets you follow them without building the infrastructure yourself.a # that's all 69 ## 相关链接 - [self.dll](https://x.com/seelffff) - [@seelffff](https://x.com/seelffff) - [12K](https://x.com/seelffff/status/2049214021430325677/analytics) - [ollama](https://github.com/ollama/ollama) - [github.com/ollama/ollama](https://github.com/ollama/ollama) - [llama.cpp](https://github.com/ggml-org/llama.cpp) - [github.com/ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) - [vllm](https://github.com/vllm-project/vllm) - [github.com/vllm-project/vllm](https://github.com/vllm-project/vllm) - [lm studio](https://github.com/lmstudio-ai/lmstudio.js) - [github.com/lmstudio-ai/lmstudio.js](https://github.com/lmstudio-ai/lmstudio.js) - [jan](https://github.com/janhq/jan) - [github.com/janhq/jan](https://github.com/janhq/jan) - [text-generation-webui](https://github.com/oobabooga/text-generation-webui) - [github.com/oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) - [localai](https://github.com/mudler/LocalAI) - [github.com/mudler/LocalAI](https://github.com/mudler/LocalAI) - [langchain](https://github.com/langchain-ai/langchain) - [github.com/langchain-ai/langchain](https://github.com/langchain-ai/langchain) - [llamaindex](https://github.com/run-llama/llama_index) - [github.com/langchain-ai/langchain](https://github.com/langchain-ai/langchain) - [rag-anything](https://github.com/HKUDS/RAG-Anything) - [github.com/HKUDS/RAG-Anything](https://github.com/HKUDS/RAG-Anything) - [chroma](https://github.com/chroma-core/chroma) - [github.com/chroma-core/chroma](https://github.com/chroma-core/chroma) - [weaviate](https://github.com/weaviate/weaviate) - [github.com/weaviate/weaviate](https://github.com/weaviate/weaviate) - [haystack](https://github.com/deepset-ai/haystack) - [github.com/deepset-ai/haystack](https://github.com/deepset-ai/haystack) - [docling](https://github.com/DS4SD/docling) - [github.com/DS4SD/docling](https://github.com/DS4SD/docling) - [autogen](https://github.com/microsoft/autogen) - [github.com/microsoft/autogen](https://github.com/microsoft/autogen) - [crewai](https://github.com/crewAIInc/crewAI) - [github.com/crewAIInc/crewAI](https://github.com/crewAIInc/crewAI) - [langgraph](https://github.com/langchain-ai/langgraph) - [github.com/langchain-ai/langgraph](https://github.com/langchain-ai/langgraph) - [agno](https://github.com/agno-agi/agno) - [github.com/agno-agi/agno](https://github.com/agno-agi/agno) - [smolagents](https://github.com/huggingface/smolagents) - [github.com/huggingface/smolagents](https://github.com/huggingface/smolagents) - [openhands](https://github.com/All-Hands-AI/OpenHands) - [github.com/All-Hands-AI/OpenHands](https://github.com/All-Hands-AI/OpenHands) - [superagi](https://github.com/TransformerOptimus/SuperAGI) - [github.com/TransformerOptimus/SuperAGI](https://github.com/TransformerOptimus/SuperAGI) - [dspy](https://github.com/stanfordnlp/dspy) - [github.com/stanfordnlp/dspy](https://github.com/stanfordnlp/dspy) - [guidance](https://github.com/guidance-ai/guidance) - [github.com/guidance-ai/guidance](https://github.com/guidance-ai/guidance) - [outlines](https://github.com/dottxt-ai/outlines) - [github.com/dottxt-ai/outlines](https://github.com/dottxt-ai/outlines) - [promptfoo](https://github.com/promptfoo/promptfoo) - [github.com/promptfoo/promptfoo](https://github.com/promptfoo/promptfoo) - [braintrust](https://github.com/brainlid/langchain) - [github.com/brainlid/langchain](https://github.com/brainlid/langchain) - [instructor](https://github.com/instructor-ai/instructor) - [github.com/instructor-ai/instructor](https://github.com/instructor-ai/instructor) - [unsloth](https://github.com/unslothai/unsloth) - [github.com/unslothai/unsloth](https://github.com/unslothai/unsloth) - [axolotl](https://github.com/axolotl-org/axolotl) - [github.com/axolotl-org/axolotl](https://github.com/axolotl-org/axolotl) - [llama-factory](https://github.com/hiyouga/LLaMA-Factory) - [github.com/hiyouga/LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) - [trl](https://github.com/huggingface/trl) - [github.com/huggingface/trl](https://github.com/huggingface/trl) - [torchtune](https://github.com/pytorch/torchtune) - [github.com/pytorch/torchtune](https://github.com/pytorch/torchtune) - [mergekit](https://github.com/arcee-ai/mergekit) - [github.com/arcee-ai/mergekit](https://github.com/arcee-ai/mergekit) - [markitdown](https://github.com/microsoft/markitdown) - [github.com/microsoft/markitdown](https://github.com/microsoft/markitdown) - [files-to-prompt](https://github.com/simonw/files-to-prompt) - [github.com/simonw/files-to-prompt](https://github.com/simonw/files-to-prompt) - [crawl4ai](https://github.com/unclecode/crawl4ai) - [github.com/unclecode/crawl4ai](https://github.com/unclecode/crawl4ai) - [firecrawl](https://github.com/mendableai/firecrawl) - [github.com/mendableai/firecrawl](https://github.com/mendableai/firecrawl) - [playwright-mcp](https://github.com/microsoft/playwright-mcp) - [github.com/microsoft/playwright-mcp](https://github.com/microsoft/playwright-mcp) - [model-context-protocol](https://github.com/anthropics/model-context-protocol) - [github.com/anthropics/model-context-protocol](https://github.com/anthropics/model-context-protocol) - [awesome-mcp-servers](https://github.com/punkpeye/awesome-mcp-servers) - [github.com/punkpeye/awesome-mcp-servers](https://github.com/punkpeye/awesome-mcp-servers) - [n8n](https://github.com/n8n-io/n8n) - [github.com/n8n-io/n8n](https://github.com/n8n-io/n8n) - [$50K](https://x.com/search?q=%2450K&src=cashtag_click) - [litellm](https://github.com/BerriAI/litellm) - [github.com/BerriAI/litellm](https://github.com/BerriAI/litellm) - [bentoml](https://github.com/bentoml/BentoML) - [github.com/bentoml/BentoML](https://github.com/bentoml/BentoML) - [ray serve](https://github.com/ray-project/ray) - [github.com/ray-project/ray](https://github.com/ray-project/ray) - [triton inference server](https://github.com/triton-inference-server/server) - [github.com/triton-inference-server/server](https://github.com/triton-inference-server/server) - [lorax](https://github.com/predibase/lorax) - [github.com/predibase/lorax](https://github.com/predibase/lorax) - [supabase](https://github.com/supabase/supabase) - [github.com/supabase/supabase](https://github.com/supabase/supabase) - [obra/superpowers](https://github.com/obra/superpowers) - [github.com/obra/superpowers](https://github.com/obra/superpowers) - [claude-code-skills](https://github.com/anthropics/claude-code-skills) - [github.com/anthropics/claude-code-skills](https://github.com/anthropics/claude-code-skills) - [free-claude-code](https://github.com/Alishahryar1/free-claude-code) - [github.com/Alishahryar1/free-claude-code](https://github.com/Alishahryar1/free-claude-code) - [claude-mem](https://github.com/thedotmack/claude-mem) - [github.com/thedotmack/claude-mem](https://github.com/thedotmack/claude-mem) - [unstructured](https://github.com/Unstructured-IO/unstructured) - [github.com/Unstructured-IO/unstructured](https://github.com/Unstructured-IO/unstructured) - [datatrove](https://github.com/huggingface/datatrove) - [github.com/huggingface/datatrove](https://github.com/huggingface/datatrove) - [trafilatura](https://github.com/adbar/trafilatura) - [github.com/adbar/trafilatura](https://github.com/adbar/trafilatura) - [semchunk](https://github.com/umarbutler/semchunk) - [github.com/umarbutler/semchunk](https://github.com/umarbutler/semchunk) - [datachain](https://github.com/iterative/datachain) - [github.com/iterative/datachain](https://github.com/iterative/datachain) - [moondream](https://github.com/vikhyat/moondream) - [github.com/vikhyat/moondream](https://github.com/vikhyat/moondream) - [internvl](https://github.com/OpenGVLab/InternVL) - [github.com/OpenGVLab/InternVL](https://github.com/OpenGVLab/InternVL) - [whisper](https://github.com/openai/whisper) - [github.com/openai/whisper](https://github.com/openai/whisper) - [insanely-fast-whisper](https://github.com/Vaibhavs10/insanely-fast-whisper) - [github.com/Vaibhavs10/insanely-fast-whisper](https://github.com/Vaibhavs10/insanely-fast-whisper) - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) - [github.com/AUTOMATIC1111/stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) - [kreo - copy the top polymarket traders](https://t.me/KreoPolyBot?start=ref-kreohub) - [http://t.me/KreoPolyBot?start=ref-kreohub](http://t.me/KreoPolyBot?start=ref-kreohub) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [3:47 AM · Apr 29, 2026](https://x.com/seelffff/status/2049214021430325677) - [12.8K Views](https://x.com/seelffff/status/2049214021430325677/analytics) - [View quotes](https://x.com/seelffff/status/2049214021430325677/quotes) --- *导出时间: 2026/4/29 13:39:28* --- ## 中文翻译 # 2026 年 4 月最佳的 69 个开源 AI 仓库 **作者**: self.dll **日期**: 2026-04-28T19:47:49.000Z **来源**: [https://x.com/seelffff/status/2049214021430325677](https://x.com/seelffff/status/2049214021430325677) --- ![Image](/doc/1264/image/img_53da7889e793.jpg) 普通的 AI 初创公司每年在工具订阅上要花费 8,000 到 50,000 美元。真正在构建 AI 的人——研究人员、工程师、独立黑客——使用的是开源。他们一直如此。我浏览了 300 多个仓库。筛选出了 69 个生产就绪、积极维护且真正有用的工具。这就是完整的 AI 技术栈。免费。 ## 这就是工具箱。里面有什么： - 01 llm inference（大模型推理） - 本地运行模型。没有 api。没有限制。没有账单。 - 02 rag & knowledge（检索增强与知识） - 让 ai 根据你自己的数据回答。 - 03 ai agents（智能体） - 让模型去行动，而不仅仅是回答。 - 04 prompts & evals（提示词与评估） - 停止猜测。开始测量。 - 05 fine-tuning（微调） - 让模型变成你自己的。 - 06 tools & context（工具与上下文） - 给模型提供它真正需要的东西。 - 07 deployment（部署） - 发布它。然后扩展它。 - 08 claude-specific（Claude 专属） - 如果你使用 claude，这些不是可选的。 - 09 data prep（数据准备） - 垃圾进，垃圾出。修复输入。 - 10 vision & multimodal（视觉与多模态） - 超越文本。 - 69 bonus（彩蛋） - 不想自己运行所有这些？ ![Image](/doc/1264/image/img_9fd583c063b5.jpg) # 01 llm inference（大模型推理）本地运行模型。没有 api。没有限制。没有账单。 ## 01 ollama ★ 98K - github.com/ollama/ollama 使用一条命令在本地运行 llama、mistral、qwen、gemma。让你的机器跑起来模型的最快方式。`ollama run llama3` 搞定。支持 gpu 加速、rest api、openai 兼容的端点。 ## 02 llama.cpp ★ 72K - github.com/ggml-org/llama.cpp 纯 c++ 编写的 llm 推理引擎。运行在 cpu、gpu、apple silicon 上。大多数本地 ai 工具背后的引擎。如果说 ollama 是汽车，llama.cpp 就是引擎。极快，低内存占用。 ## 03 vllm ★ 44K - github.com/vllm-project/vllm 生产环境的高吞吐量 llm 服务引擎。连续批处理、分页注意力、openai 兼容 api。大规模部署模型的标准。大多数正经 ai 公司都在用。 ## 04 lm studio ★ 28K 带有清爽 UI 的桌面应用，用于运行本地 llm。从 hugging face 下载模型，在本地运行，获得一个 openai 兼容的本地服务器。非开发人员的最佳入门工具。 github.com/lmstudio-ai/lmstudio.js ## 05 jan ★ 26K - github.com/janhq/jan 开源的 chatgpt 替代品，100% 离线运行。UI 整洁、模型管理、本地 api 服务器。适用于 mac、windows、linux。没有数据会离开你的机器。 ## 06 text-generation-webui ★ 42K - github.com/oobabooga/text-generation-webui 本地 llm 的瑞士军刀。支持每一种模型格式、每一个后端、每一个采样器。角色模式、笔记本模式、api 模式。目前功能最全的本地 UI。 ## 07 localai ★ 26K - github.com/mudler/LocalAI 自托管的 openai 直接替代品。同样的 api，本地模型。在任何应用中将 claude/gpt 替换为本地 llm，无需更改一行代码。 # 02 rag & knowledge（检索增强与知识）让 ai 根据你自己的数据回答。 ## 08 langchain ★ 98K - github.com/langchain-ai/langchain 最流行的 llm 框架。链、智能体、检索器、记忆。将 llm 连接到任何数据源或工具。庞大的集成生态系统。如果你用 ai 构建任何东西，你最终都会遇到 langchain。 ## 09 llamaindex ★ 38K - github.com/langchain-ai/langchain llm 应用的数据框架。索引任何数据源——pdf、sql、notion、slack——并用自然语言查询。对于纯 rag 用例来说比 langchain 更好。 ## 10 rag-anything ★ 12K - github.com/HKUDS/RAG-Anything 面向 claude 和其他 llm 的多模态 rag。处理文本、表格、图片、图表、图形。不仅仅是 pdf——实际上是所有东西。6 行代码即可设置。正经的团队在生产中使用。 ## 11 chroma ★ 16K - github.com/chroma-core/chroma 开源向量数据库。存储嵌入，按相似度搜索，按元数据过滤。可以运行在内存中或持久化。给任何项目添加语义搜索的最简单方法。 ## 12 weaviate ★ 12K - github.com/weaviate/weaviate 内置机器学习模型的向量数据库。混合搜索、多租户、实时更新。生产就绪，可扩展至数十亿对象。你听说过的公司都在用。 ## 13 haystack ★ 18K - github.com/deepset-ai/haystack 用于 rag 流水线的端到端 nlp 框架。模块化、生产就绪，适用于任何 llm 或向量数据库。目前最成熟的 rag 框架。 ## 14 docling ★ 22K - github.com/DS4SD/docling 将文档转换为结构化 markdown 以供 ai 使用。处理带有表格、图形、公式的 pdf——不仅仅是纯文本提取。由 ibm research 构建。 # 03 ai agents（智能体）让模型去行动，而不仅仅是回答。 ## 15 autogen ★ 40K - github.com/microsoft/autogen 微软的多智能体对话框架。智能体互相交谈、委派任务、编写并执行代码。用于复杂智能体工作流的最强大框架。 ## 16 crewai ★ 28K - github.com/crewAIInc/crewAI 编排角色扮演 ai 智能体。定义一组人马、分配角色、设定目标——智能体像团队一样协作。构建真正有效的多智能体系统的最简单方法。 ## 17 langgraph ★ 10K - github.com/langchain-ai/langgraph 将有状态的多智能体工作流构建为图。节点是智能体或函数，边是转换。处理复杂逻辑、循环、人在回路。生产级智能体框架。 ## 18 agno ★ 22K - github.com/agno-agi/agno 构建快速的多模态 ai 智能体。支持任何 llm、任何工具、记忆、知识、存储。对于简单智能体，比 langchain 快 10 倍。API 整洁，文档优秀。 ## 19 smolagents ★ 14K - github.com/huggingface/smolagents hugging face 的极简智能体框架。编写并执行 python 代码以解决任务的代码智能体。非常简单——总共只有 1000 行代码。反 langchain 主义者。 ## 20 openhands ★ 48K - github.com/All-Hands-AI/OpenHands 开源的 devin 替代品。ai 软件工程师，编写代码、运行测试、修复 bug、进行部署。适用于 claude、gpt-4、本地模型。能力最强的编码智能体。 ## 21 superagi ★ 16K - github.com/TransformerOptimus/SuperAGI 自托管自主 ai 智能体基础设施。智能体市场、性能遥测、并发智能体、图形 UI。在你自己的服务器上并行运行多个智能体。 ## 04 prompts & evals（提示词与评估）停止猜测。开始测量。 ## 29 dspy ★ 22K - github.com/stanfordnlp/dspy 编程——而不是提示——llm。定义你想要什么，dspy 会自动优化提示。来自 stanford nlp。用系统优化取代手动提示工程。 ## 30 guidance ★ 20K - github.com/guidance-ai/guidance 用代码控制 llm 输出结构。将生成与逻辑交错，强制 json 模式，约束输出。当你需要模型准确输出你需要的东西时。 ## 31 outlines ★ 11K - github.com/dottxt-ai/outlines 结构化文本生成。强制模型输出有效的 json、正则模式、特定格式。零提示工程——保证输出结构。 32 promptfoo ★ 6K - github.com/promptfoo/promptfoo 测试并评估你的提示。运行自动化测试，比较模型，捕获回归。就像单元测试，但是针对 ai。在将任何东西发布到生产环境之前必不可少。 ## 33 braintrust ★ 3K - github.com/brainlid/langchain llm 应用的评估框架。跨模型版本、提示和配置跟踪质量。因为直觉不是一种指标。 ## 34 instructor ★ 9K - github.com/instructor-ai/instructor 使用 pydantic 从 llm 获取结构化输出。定义一个模式，取回一个经过验证的 python 对象。适用于 openai、anthropic、google、本地模型。最整洁的结构化输出解决方案。 # 05 fine-tuning（微调）让模型变成你自己的。 ## 35 unsloth ★ 24K - github.com/unslothai/unsloth 以 2 倍速度微调 llm，内存减少 80%。支持 llama、mistral、qwen、gemma。在单个 gpu 上运行。如果你资源受限，这是你唯一需要的微调库。 ## 36 axolotl ★ 8K - github.com/axolotl-org/axolotl 精简的 llm 微调工具。yaml 配置、每一种数据集格式、每一种训练技术。hugging face transformers 之上的运维层。大多数正经微调人员都在用。 ## 37 llama-factory ★ 40K - github.com/hiyouga/LLaMA-Factory 零代码微调 100+ 个 llm。Web UI，支持 lora、qlora、全量微调。目前最用户友好的微调工具。40K 星不是没道理的。 ## 38 trl ★ 12K - github.com/huggingface/trl transformer 强化学习。rlhf、dpo、ppo——用于对齐 gpt-4 和 claude 的技术。来自 hugging face。当你想要训练模型真正做你想做的事时使用。 ## 39 torchtune ★ 5K - github.com/pytorch/torchtune 来自 meta 的 pytorch 原生微调库。简单、可黑入、文档完善。纯 pytorch 微调的参考实现。 ## 40 mergekit ★ 4K - github.com/arcee-ai/mergekit 将多个微调后的模型合并为一个。slerp、ties、dare、线性合并——所有技术。合并无需 gpu。创造性能优于父模型的弗兰肯斯坦模型。每个在 hugging face 上发布合并模型的人都在用。 # 06 tools & context（工具与上下文）给模型提供它真正需要的东西 ## 41 markitdown ★ 38K - github.com/microsoft/markitdown 将任何文件转换为 markdown。pdf、word、excel、powerpoint、图片、音频。向你的 llm 输送整洁的结构化文本，而不是垃圾。来自 microsoft。38K 星。 ## 42 files-to-prompt ★ 3K - github.com/simonw/files-to-prompt 将你的整个代码库变成一个提示。尊重 .gitignore、递归、可过滤。来自 simon willison。将项目输送给 claude 的最简单工具。 ## 43 crawl4ai ★ 30K - github.com/unclecode/crawl4ai 面向 ai 的网页抓取。从任何 url 提取整洁的 markdown，处理重度 js 的站点、结构化数据提取。任何 ai 流水线的 Web 数据层。 ## 44 firecrawl ★ 25K - github.com/mendableai/firecrawl 将任何网站转换为 llm 就绪的数据。全站爬取、结构化提取、整洁的 markdown 输出。面向 ai 应用的生产级网页爬虫。 ## 45 playwright-mcp ★ 31K - github.com/microsoft/playwright-mcp 给 claude 一个真正的浏览器。导航、点击、截图、阅读动态内容。在 30 秒内分析任何站点。面向 web 任务的最强大的 mcp 服务器。 ## 46 model-context-protocol ★ 11K - github.com/anthropics/model-context-protocol 将 claude 连接到外部工具的标准。官方 anthropic mcp。插入任何 api、数据库、服务。生态系统中有数百个服务器。 ## 47 awesome-mcp-servers ★ 27K - github.com/punkpeye/awesome-mcp-servers 500+ 个现成的 mcp 服务器。github、slack、notion、数据库、浏览器、金融。在一个目录中集成你需要的所有东西。 ## 48 n8n ★ 47K - github.com/n8n-io/n8n 具有 400+ 集成的自托管工作流自动化。将 llm 连接到任何应用。根据计划或 webhook 触发 ai 工作流。在节点中运行自定义 js/python。正经 ai 流水线背后的自动化层。取代 50K 美元/年的 zapier。 # 07 deployment（部署）发布它。然后扩展它。 ## 49 litellm ★ 16K - github.com/BerriAI/litellm 一个 api 对接 100+ 个 llm。openai 格式，适用于 claude、gpt、gemini、本地模型。负载均衡、故障转移、成本跟踪。你的应用与每个 llm 提供商之间的代理层。 ## 50 bentoml ★ 7K - github.com/bentoml/BentoML 构建和部署 ai 服务。打包模型、创建 api、部署到任何地方。从本地测试到生产 kubernetes。不需要 devops 团队的 mlops 层。 ## 51 ray serve ★ 34K - github.com/ray-project/ray 大规模分布式 ai 推理。服务多个模型、自动扩展、处理数百万请求。openai、anyscale、生产 ai 公司在使用。大材小用，直到你需要它。 ## 52 triton inference server ★ 8K - github.com/triton-inference-server/server nvidia 的生产推理服务器。最大的 gpu 利用率、动态批处理、多模型服务。企业级 gpu 推理的标准。 ## 53 lorax ★ 3K - github.com/predibase/lorax 在一个 gpu 上服务数百个 lora 微调模型。一个基础模型，动态加载数百个适配器。服务微调模型的成本降低 10 倍。 ## 54 supabase ★ 73K - github.com/supabase/supabase ai 应用的默认后端。基于 postgres 构建的开源 firebase 替代品。实时数据库、认证、存储、边缘函数、向量搜索。73K 星。取代 firebase + auth0（15K 美元/年）。 # 08 claude-specific（Claude 专属）如果你使用 claude，这些不是可选的。 ## 55 obra/superpowers ★ 160K - github.com/obra/superpowers 为 claude code 添加超能力。深度代码分析、自动重构、项目级编辑。作为官方 cli 之上的层工作。160K 星。最受欢迎的 claude 增强。 ## 56 claude-code-skills ★ 官方 - github.com/anthropics/claude-code-skills 官方 anthropic 技能框架。skill.md 模式，教导 claude 无误地处理文档、自动化、工作流。claude code 处理复杂任务的基础。 ## 57 free-claude-code ★ 2K - github.com/Alishahryar1/free-claude-code 通过 github models api 完全免费运行 claude code。在 github 上趋势排名第一。分步指南 + 设置脚本。0 美元。永远。 ## 58 claude-mem ★ 1K - github.com/thedotmack/claude-mem claude 的持久记忆。自动捕获 claude 在跨会话中所做的一切。取代付费的上下文管理工具。claude 记得你是谁以及你在做什么。 # 09 data prep（数据准备）垃圾进，垃圾出。修复输入。 ## 59 unstructured ★ 10K - github.com/Unstructured-IO/unstructured 为 llm 提取和转换非结构化数据。pdf、html、word、图片、邮件——全部解析为整洁的块，准备用于 rag。大多数 ai 流水线缺少的数据层。 ## 60 datatrove ★ 3K - github.com/huggingface/datatrove 面向 llm 训练的大规模数据处理。来自 hugging face。处理 tb 级文本，具有去重、质量过滤和内容分类。大实验室使用的工具。 ## 61 trafilatura ★ 3K -github.com/adbar/trafilatura 面向 ai 的网页内容提取。去除样板文件，保留内容，输出整洁的文本或 markdown。用于向模型输送文本的最佳单页提取器。 62 semchunk ★ 1K - github.com/umarbutler/semchunk 面向 rag 的语义文本分块。在自然边界而不是任意 token 计数处分割文本。更好的块 → 更好的检索 → 更好的答案。 ## 63 datachain ★ 2K - github.com/iterative/datachain ai 原生数据集管理。版本控制、查询和转换多模态数据集。适用于图片、视频、文本、嵌入。为 llm 训练工作流构建 # 10 vision & multimodal（视觉与多模态）超越文本。 ## 64 moondream ★ 10K - github.com/vikhyat/moondream 可在任何地方运行的微型视觉语言模型。16 亿参数。描述图片、回答视觉问题、检测物体。在树莓派上运行。最小的有用的视觉模型。 ## 65 internvl ★ 7K - github.com/OpenGVLab/InternVL 最先进的开源视觉模型。在大多数基准测试中与 gpt-4v 匹配。理解图片、图表、文档、截图。claude 视觉功能的开放替代品。 ## 66 whisper ★ 74K - github.com/openai/whisper openai 的开源语音识别。转录 99 种语言的音频。本地运行，处理口音、背景噪音、技术术语。将音频输送给你的 llm 流水线。 ## 67 insanely-fast-whisper ★ 8K - github.com/Vaibhavs10/insanely-fast-whisper whisper 但快 10-20 倍。一个命令、自动 gpu 优化、批处理。在消费级硬件上 2 分钟内转录 2 小时的播客。 ## 68 stable-diffusion-webui ★ 143K - github.com/AUTOMATIC1111/stable-diffusion-webui stable diffusion 的浏览器界面。从文本生成、编辑、放大图片。143K 星——除了 llm 之外星标最多的 ai 仓库。数百个扩展、风格、controlnet、重绘。在你的 gpu 上运行。 # 69 kreo - copy the top polymarket traders ★ - http://

🔗 相关文档