The basics
An open source AI modelis a machine-learning model whose weights, architecture and training code are released under a license that lets you download, inspect, modify and redistribute the system. That distinguishes them from closed frontier models — which you can only reach through a vendor's API — and lets you run them on your own hardware, fine-tune them on your own data, and ship them inside your product without sending every request to a third party.
People often conflate open source with open weight, but they are not the same thing. An open-source release publishes the model weights plus everything you would need to reproduce them: training data, training scripts, the evaluation harness and a license. An open-weight release shares the weights but keeps some of those pieces private. Both let you self-host; only the former lets you fully reproduce the result.
Not always — “free to download” and “free for commercial use” are different questions. Permissive licenses such as Apache 2.0 and MIT let you use the model in a commercial product, modify it and redistribute it with very few conditions. Others, like some community or research licenses, restrict commercial use above a revenue threshold or forbid using outputs to train competing models. Before you ship, check the License and Commercial use filters on each model card and read the original model card — that is the only source that is legally binding.
You do not need a data centre to start. For laptops and single-GPU desktops, Ollama and llama.cpp run quantised models (4-bit and 8-bit builds) with a one-line install, trading a little accuracy for a much smaller memory footprint. For production serving and high throughput, vLLM and Text Generation Inference batch requests across a GPU efficiently. As a rough guide: a 7B model needs roughly 4–6 GB of VRAM when quantised, a 13B model around 8–10 GB, and 70B-class models realistically want multiple GPUs or aggressive quantisation. Use the Parameters and Context window filters above to find a model that fits your hardware.
Hosted API models are fastest to integrate and need zero infrastructure, but you pay per token, your data leaves your network, and you cannot pin a specific version forever. Self-hosted open models flip every one of those trade-offs: fixed infrastructure cost instead of per-request billing, full data privacy, complete version control and the freedom to fine-tune. A common pattern is to prototype against a hosted API, then move the steady, high-volume workloads to an open model you host yourself once the economics make sense. Many open models are also available through hosted providers (Together AI, Groq, Fireworks, Hugging Face), so you can get both without rewriting your code.
Browse
FAQ
There is no single best model — it depends on the task. For general chat and reasoning, large Llama, Mistral and Qwen models lead the open leaderboards; for code, models like DeepSeek-Coder and Code Llama are strong; for images, the Stable Diffusion family; and for speech, Whisper. Use the task tabs and sort options above to compare the current top models by downloads, parameters and license.
Yes. Small and quantised models (1B–7B parameters in 4-bit form) run on a modern CPU with tools like llama.cpp and Ollama, though responses are slower than on a GPU. Embedding models and many speech models are also comfortable on CPU. For larger models or real-time use, a consumer GPU with 8GB or more of VRAM makes a noticeable difference.
They can be, provided you treat them like any other dependency: pin a specific version, review the license, scan the weights file for tampering, and evaluate the model on your own test set before launch. Because you host the model yourself, no prompt or user data is sent to a third party, which is often easier to reconcile with privacy and compliance requirements than a closed API.
An open-weight model publishes its trained weights so you can download and run it, but may keep the training data, training code or full license private. An open-source model publishes the weights and everything needed to reproduce them. Both can be self-hosted; only a fully open-source release lets you independently retrain or audit how the model was built.
The model itself is free to download. Your real cost is compute: a single cloud GPU instance suitable for a 7B–13B model typically runs a few hundred dollars a month if left on continuously, or a fraction of that if you autoscale to zero between requests. Self-hosting becomes cheaper than a per-token API once your request volume is high and steady.
Every model listed here is curated from public releases by their original developers — companies and labs such as Meta, Mistral, Google, Alibaba and independent research groups. Each card links to the official model card and repository so you can verify the license, parameters and capabilities at the source.