Top 20 libraries for building with AI

The engine under most local AI apps; runs quantized models on CPUs, Macs and GPUs.

Best for Developers building local AI apps that must run on many kinds of hardwareYou give up Low level; Ollama and LM Studio run the same models with an easier setup

Apple's framework for running and training models on Apple Silicon Macs.

Best for Mac developers running or training models on Apple SiliconYou give up Apple Silicon only; slows down at long context in September tests

Widely used framework for LLM apps and agents, with integrations for almost every provider.

Best for Developers who need ready links to almost every model providerYou give up Often criticised as heavy, and simpler options exist; not scored yet

TypeScript toolkit for calling any model, streaming chat UIs and building agents in JS apps.

Best for JavaScript and TypeScript developers adding streaming chat to web appsYou give up Built for JavaScript apps; not scored: no independent test in 90 days

The standard Python library for loading, running and fine-tuning nearly every open model.

Best for Python developers loading, running or fine-tuning almost any open modelYou give up A library for developers, not an end tool; not scored yet

Security warning: 1.107.0 (21 Oct) was malware; use 1.106.2, rotate keys. One API for 100+ providers

Best for Teams routing calls to many model providers with cost tracking and fallbacksYou give up Poisoned releases have hit PyPI more than once, latest 21 Oct; pin versions

The base framework most model training and inference code is written in.

Best for Researchers and engineers writing their own training or model codeYou give up A base framework, not a ready tool; not scored: no recent independent test

Graph-based framework for agents that need state, retries and human sign-off steps.

Best for Developers who want tight control over agents in productionYou give up Less quick to start than CrewAI; not scored yet

AWS's model-driven agent SDK for Python and TypeScript; common in Bedrock projects.

Best for Teams building agents in AWS Bedrock projects, in Python or TypeScriptYou give up Most at home in AWS projects; not scored: no recent independent test

Python agent framework from the Pydantic team, built around typed, checkable outputs.

Best for Python developers who want typed, testable agent outputsYou give up Far fewer downloads than LangChain; not scored yet

OpenAI's lightweight framework for multi-agent apps with handoffs and guardrails.

Best for Developers building multi-agent apps with handoffs, guardrails and tracingYou give up Still before version 1.0; not scored: no independent evaluation in 90 days

Google's agent framework for Python and JavaScript; works with Gemini and other models.

Best for Developers building agents on Gemini who may switch models laterYou give up Tuned for Gemini first; not scored: no recent independent test

Python library that lets an AI model drive a real web browser to finish tasks on websites.

Best for Python developers automating tasks on real websitesYou give up Cloud bills add model cost on top of $0.02 an hour; not scored yet

Framework for connecting AI models to your documents; strong on retrieval (RAG).

Best for Developers connecting models to their own documents and parsing filesYou give up Its document parser, LlamaParse, costs from $50 a month; not scored yet

Python library for running and training image, video and audio diffusion models.

Best for Python developers running or training image, video and audio modelsYou give up A library for developers, not an end tool; not scored yet

Fast GPU serving engine; a strong pick for agent workloads that reuse long shared prompts.

Best for Serving large open models on graphics cards for agent workYou give up Only older secondhand benchmarks exist; not scored until a new test appears

Python framework for teams of role-based agents; quick to start, less control than LangGraph.

Best for Python developers who want a team of role-based agents running fastYou give up Far fewer downloads than LangGraph; hosted enterprise plan is by quote

The usual choice for serving open models to many users at once on data center GPUs.

Best for Serving open models to many users at once in productionYou give up Only older benchmarks exist, so not scored; load only models you trust

Memory layer that lets agents and chat apps remember users across sessions.

Best for Developers adding long-term user memory to an agent or chat appYou give up Hosted plan from $19 a month; not scored: no recent independent test

Fine-tune open models on one GPU with less memory; multi-GPU support is weaker.

Best for Fine-tuning open models on one graphics card, now with a local appYou give up Speed claims come only from Unsloth's own figures; Pro is by quote