DeepSeek-R1
DeepSeek · Language model
The reasoning model that showed open weights could match frontier labs on maths and code, at a fraction of the training cost.
Open-source AI
Dify
LangGenius · Agent framework
A visual platform to build chatbots, agents and RAG apps without much code, and self-host the result.
FLUX.1
Black Forest Labs · Image, video & audio
Image models from the original Stable Diffusion researchers. Schnell is fully open; Dev is free for non-commercial use.
gpt-oss
OpenAI · Language model
OpenAI’s first open-weight reasoning models since GPT-2. The smaller one fits on a single high-end consumer GPU.
LangGraph
LangChain · Agent framework
Build agents as graphs with explicit steps, loops and human approval points. Popular for production agents.
Ollama
Ollama · Runtime & tooling
The easiest way to run open models on your own machine. One command to download, one to chat.
OpenHands
All Hands AI · AI agent
A software engineering agent that writes code, runs commands and browses docs inside a sandbox, like a junior developer.
Qwen3
Alibaba · Language model
A family that runs from phone-sized to very large, with a switch between quick answers and step-by-step thinking.
Whisper
OpenAI · Image, video & audio
Speech recognition that handles accents, noise and dozens of languages. Still the default for transcription projects.
Aider
Aider · AI agent
Pair programming in the terminal. It edits files in your git repo and commits each change with a sensible message.
AutoGen
Microsoft · Agent framework
Microsoft’s framework for agents that talk to each other to solve problems, with a visual studio for prototyping.
AutoGPT
Significant Gravitas · AI agent
One of the first autonomous agent projects, now a platform for building and running agents on a schedule.
Browser Use
Browser Use · AI agent
Lets an AI drive a real web browser: clicking, typing and reading pages to finish tasks on sites without an API.
Cline
Cline · AI agent
A coding agent inside VS Code that asks before every file edit or terminal command, so you stay in control.
Codex CLI
OpenAI · AI agent
OpenAI’s coding agent for the terminal. It reads your project, proposes changes and runs them in a sandbox.
ComfyUI
Comfy Org · Runtime & tooling
A node-based studio for image and video models. Steep at first, unmatched once you build your own workflows.
Continue
Continue · AI agent
An open assistant for VS Code and JetBrains that you can wire to local or hosted models.
CrewAI
CrewAI · Agent framework
Define a crew of role-playing agents, give each a goal, and let them hand work to each other.
DeepSeek-V3
DeepSeek · Language model
A large mixture-of-experts chat model that only activates a slice of its parameters per token, which keeps serving costs low.
Flowise
FlowiseAI · Agent framework
Drag-and-drop builder for LLM apps and agents, built on LangChain components.
Gemini CLI
Google · AI agent
Google’s open terminal agent with a generous free tier and Google Search built in.
Gemma
Google DeepMind · Language model
Google’s lightweight open models, built from Gemini research and tuned to run well on laptops and phones.
GLM-4.5
Z.ai · Language model
Z.ai’s agent-focused model with a lighter Air variant, strong at coding and web browsing tasks.
goose
Block · AI agent
A local, extensible agent from Block that connects to tools through MCP and works with any model you choose.
Haystack
deepset · Agent framework
A mature framework for search and question answering pipelines, popular with enterprise teams.
HunyuanVideo
Tencent · Image, video & audio
Tencent’s large video model, known for realistic people and smooth camera movement.
Kimi K2
Moonshot AI · Language model
A trillion-parameter mixture-of-experts model tuned for agents: it is at its best when chaining tools to finish a task.
Kokoro
hexgrad · Image, video & audio
A small text-to-speech model with surprisingly natural voices. Fast enough to run on a CPU.
LangChain
LangChain · Agent framework
The best-known toolkit for connecting language models to data, tools and memory. Huge number of integrations.
Langflow
Langflow · Agent framework
A visual editor for agent and RAG flows that you can export as an API in one click.
Llama
Meta · Language model
The model family that made open weights mainstream. Huge ecosystem of fine-tunes, quantisations and tutorials.
llama.cpp
ggml · Runtime & tooling
The C++ engine behind most local AI apps. Squeezes large models onto laptops with clever quantisation.
LlamaIndex
LlamaIndex · Agent framework
Focused on getting your documents into an AI’s answers: parsing, indexing and retrieval done well.
LocalAI
LocalAI · Runtime & tooling
A drop-in replacement for the OpenAI API that runs text, image and audio models on your own hardware.
LTX-Video
Lightricks · Image, video & audio
A video model built for speed, able to render short clips faster than they play back on strong hardware.
Mastra
Mastra · Agent framework
A TypeScript framework for agents and workflows, with memory, evals and a local playground included.
MetaGPT
FoundationAgents · AI agent
Simulates a small software company: product manager, architect and engineers as agents working from one brief.
MiniMax-M1
MiniMax · Language model
An open reasoning model built for very long inputs, using a hybrid attention design to keep long prompts affordable.
Mistral
Mistral AI · Language model
Compact European models with a reputation for punching above their size, plus a clean reference inference library.
Mochi 1
Genmo · Image, video & audio
An open video model with strong prompt adherence and natural motion, released with its full weights.
n8n
n8n · Agent framework
Workflow automation you can self-host, with AI agent nodes that plug into hundreds of apps.
OLMo
Ai2 · Language model
Truly open: the weights, training data, code and logs are all public, which makes it the go-to model for research.
Open Interpreter
Open Interpreter · AI agent
Talk to your computer in plain language and it writes and runs the code to get things done locally.
Open WebUI
Open WebUI · Runtime & tooling
A polished, ChatGPT-style interface for your own models, with users, documents and web search.
OpenAI Agents SDK
OpenAI · Agent framework
A small, readable SDK for agents with handoffs, guardrails and tracing. Works with other providers too.
Phi
Microsoft · Language model
Small language models trained on carefully chosen data. A good pick when you need decent reasoning on modest hardware.
Pydantic AI
Pydantic · Agent framework
Type-safe agents from the team behind Pydantic. Feels natural if you already write modern Python.
Qwen3-Coder
Alibaba · Language model
Qwen’s coding specialist, trained for agentic work like editing repos and calling tools over long sessions.
Segment Anything 2
Meta · Image, video & audio
Click on any object and it is cut out, in photos and across video frames. Widely used in editing and labelling tools.
smolagents
Hugging Face · Agent framework
A minimal agent library where agents write their actions as Python code, which often beats JSON tool calls.
SmolLM
Hugging Face · Language model
Tiny models that run in a browser tab or on a phone, with the full training recipe published.
Stable Diffusion 3.5
Stability AI · Image, video & audio
The latest of the models that started the open image boom, with a huge library of community styles and add-ons.
SWE-agent
Princeton & Stanford · AI agent
A research agent that fixes real GitHub issues and helped define how coding agents are benchmarked.
Transformers
Hugging Face · Runtime & tooling
The library that loads almost every open model with a few lines of Python. Foundation of the open AI ecosystem.
Ultralytics YOLO
Ultralytics · Image, video & audio
Real-time object detection that you can train on your own images in an afternoon.
vLLM
vLLM project · Runtime & tooling
High-throughput serving for language models. The standard choice when you host open models for many users.
Wan
Alibaba · Image, video & audio
Open video generation that rivals paid services for motion and detail, and runs on a single consumer GPU at lower resolutions.
Choosing a project
Start with the licence, not the benchmark. A model that tops a chart is no use if its terms rule out your product. Permissive projects (MIT and Apache 2.0) are the safest bet for commercial work. Copyleft projects are fine for internal tools, but anything you distribute must stay open. Custom terms vary a lot, from a simple attribution rule to limits on very large companies, so read them before you commit.
Next, match the size to your hardware. Small models such as Phi, Gemma and SmolLM run on an ordinary laptop. Mid-sized ones want a recent gaming GPU, and the largest mixture-of-experts models are really meant for servers. Tools like Ollama and llama.cpp make the first try painless, and vLLM is what most teams reach for once real users arrive.
For agents, decide how much control you need. Cline and Aider check with you before each change. OpenHands and Codex CLI work more independently inside a sandbox. If you are building your own, LangGraph suits careful production flows, CrewAI is quick for multi-agent ideas, and Dify or Flowise get you there without much code. Prefer hosted tools? See our Top 100 AI tools and the AI leaderboard.
FAQ
What is the difference between open source and open weights?
Open source usually means the code and model can be used, changed and shared under a recognised licence such as MIT or Apache 2.0. Open weights means you can download and run the model, but the licence may limit commercial use or very large deployments. Each card on this page says which one applies.
Which open-source model should I start with?
For general chat on your own computer, try Qwen3 or Gemma through Ollama. For coding, Qwen3-Coder or gpt-oss. If your laptop is modest, Phi and SmolLM run comfortably on a CPU.
Can I use these in a commercial product?
Projects marked Permissive generally allow it. Copyleft licences require you to share changes you distribute, and Custom terms need a closer read. Always check the licence file in the GitHub repo before you ship.
How do the votes work?
Anyone can upvote a project once. Votes are stored on HyzenPro and shared with every visitor, and repeated votes from one network are capped, so the counts reflect real readers.