AI & Technology
50 open-source projects and open-weight models you can download and run on your own machine — no subscription, no usage meter, and nothing of yours uploaded to somebody else’s server.
Every license below was read from the project’s own license file or model card on 22 September 2026, and the badge links straight to that text so you never have to take my word for it. That matters more than it sounds: several of the best-known “open” models are free to download but are not open source, and their terms decide whether you can use them for paid client work. Licenses change — check the link before you build a business on one.
Want the hosted versions instead?
If you would rather not install anything, the tools directory covers free and free-tier services you use in a browser, and the models page compares what each free tier actually gets you.
These are the programs that load a model and talk to it. Start here: pick one, then pick weights from the next section.
Ollama
Downloads and runs open models with one command, and serves them to other programs on your machine.
What you need to run it: A normal laptop will run a small model. Around 8GB of memory for a 7–8 billion parameter model; more memory means bigger models, a graphics card means faster ones.
ggml / Georgi Gerganov
The engine underneath most local AI apps. Runs models in a compressed format on ordinary hardware, including old machines.
What you need to run it: Comfortable with a terminal. You compile it, or download a release build, then point it at a model file.
Ettore Di Giacinto
A drop-in replacement for the OpenAI API that runs on your own hardware, covering text, images, speech and vision.
What you need to run it: Docker, or a single binary. No graphics card required, though one helps.
Menlo Research
A desktop app that looks like a chat app and runs models entirely offline on your machine.
What you need to run it: Download and install, like any other app. No account, no terminal.
Nomic AI
Desktop app for chatting with local models, including over your own folder of documents.
What you need to run it: An installer for Windows, Mac or Linux. Works without a graphics card.
Open WebUI Inc.
A polished browser interface over Ollama or any OpenAI-compatible API, with users, chat history and document search.
What you need to run it: Docker or Python, and a model source such as Ollama running alongside it.
Read the license first: Not a standard open-source license. It is BSD-3 with an added clause about keeping the Open WebUI branding unless you have 50 or fewer users or a separate agreement. Fine for personal and small-team use; read it before you rebrand it for a client.
Mintplex Labs
A local-first workspace that answers questions over your own documents, with agents and multiple users.
What you need to run it: Desktop app or Docker. Bring your own local or hosted model.
Danny Avila
A self-hosted chat interface that puts local and hosted models side by side in one place.
What you need to run it: Docker and a little configuration. Best if you already host things.
oobabooga
A desktop app and web interface for local models, with tool calling and an OpenAI-compatible API.
What you need to run it: A one-click installer, then a model file. A graphics card makes it far more pleasant.
Read the license first: Strong copyleft. Using it yourself is unrestricted, but if you modify it and let other people use it over a network, you have to publish your changes.
The models themselves. All of these are free to download and run offline — but read the license column, because 'open weights' and 'open source' are not the same thing.
OpenAI
OpenAI's open-weight model, released for anyone to download and run — including commercially.
What you need to run it: Roughly 16GB of memory in its compressed form. Runs in Ollama or llama.cpp without special setup.
Alibaba Cloud (Qwen team)
A family of models from tiny to very large, strong on reasoning and on languages other than English.
What you need to run it: The 8-billion version runs on a laptop with about 8GB free. Larger sizes need a GPU.
Mistral AI
A small, fast, open model that still holds up for summarizing, drafting and classification.
What you need to run it: About 5GB compressed. One of the easiest first models to run.
Microsoft
A small model trained on carefully filtered data, aimed at reasoning well for its size.
What you need to run it: Runs on a laptop. No GPU needed for the compressed version.
Hugging Face
A tiny model published with its training recipe, so you can see how it was made, not just use it.
What you need to run it: Runs on modest hardware, including some phones and single-board computers.
Allen Institute for AI
A fully open model: weights, training data, code and checkpoints all published. The one to use if you need to know what went in.
What you need to run it: A GPU for the 7-billion version, or run a compressed build locally.
Meta
Meta's large open-weight model, widely used as the baseline for self-hosted work.
What you need to run it: Serious hardware: a multi-GPU machine, or a heavily compressed build with a lot of system memory. Not a laptop model.
Read the license first: Free, but not open source. Meta's license adds an acceptable-use policy, a requirement to credit Llama in products built on it, and a separate license if your product passes 700 million monthly users. You also have to accept the terms on the model page before downloading.
Google DeepMind
Small, efficient models that handle images as well as text — useful on one machine.
What you need to run it: The 4-billion version runs on a laptop; the larger sizes want a GPU.
Read the license first: Free to use commercially, but governed by Google's own terms and a prohibited-use policy rather than an open-source license, and Google can update that policy. You must pass the same restrictions on to anyone you give the model to.
DeepSeek
A very large, strong reasoning model, released with open weights.
What you need to run it: Server-class hardware to run yourself. Most people use it through a hosted provider.
Read the license first: Two different licenses in one repository. The code is MIT; the weights carry DeepSeek's own agreement with use restrictions attached. If you are shipping something built on the weights, read that second file, not the first.
Picture and video generation you run yourself, with no per-image charge and no upload of your work to someone else's server.
Comfy Org
A node-based workspace for image and video models. The most capable free option, and the least beginner-friendly.
What you need to run it: A graphics card with 8GB or more for comfortable use. Expect a learning curve.
Read the license first: Copyleft: if you distribute a modified version, your changes have to be published under the same license.
Invoke
A more guided image-generation studio, with a proper canvas for editing rather than a node graph.
What you need to run it: An installer, plus a graphics card for reasonable speed.
AUTOMATIC1111
The long-standing browser interface for Stable Diffusion models, with an enormous library of extensions.
What you need to run it: A graphics card, and patience with the setup.
Read the license first: Strong copyleft: offering a modified version as a service means publishing your changes.
Black Forest Labs
A fast, high-quality image model whose weights are open, unlike its bigger siblings.
What you need to run it: A graphics card with around 12GB, or a compressed build on less.
Read the license first: Only the [schnell] version is Apache-2.0. The [dev] version is non-commercial and [pro] is API-only — check which one you downloaded before using an image for paid work.
Stability AI
The widely supported image model that most free tools and tutorials are built around.
What you need to run it: A graphics card with 8GB or more. Runs in ComfyUI, InvokeAI or diffusers.
Read the license first: A responsible-AI license, not a standard open-source one: it lists uses you agree not to put the model to, and you must pass those restrictions on to anyone you share it with.
Hugging Face
The Python library for running image, video and audio generation models in your own code.
What you need to run it: Python, and a GPU for anything beyond small tests.
Transcription and speech you run locally, which matters when the recording is a client call, an interview or anything you were not given permission to upload.
OpenAI
The reference speech-to-text model: transcribes and translates audio in a long list of languages.
What you need to run it: Python. The small models run on a laptop CPU; the large one wants a GPU.
ggml / Georgi Gerganov
Whisper rewritten to run fast on ordinary machines, including phones and older laptops.
What you need to run it: A terminal and a model file. No Python, no GPU.
SYSTRAN
A much faster Whisper that uses far less memory — the practical choice for batches of recordings.
What you need to run it: Python, and a GPU if you want the full speed benefit.
Max Bain
Whisper with accurate word-level timings and speaker separation — who said what, and exactly when.
What you need to run it: Python and a GPU. Speaker separation needs an extra model you accept terms for.
Rhasspy
Fast local text-to-speech with a wide set of voices, good enough for narration and accessibility use.
What you need to run it: A small download. Runs on a Raspberry Pi, never mind a laptop.
Coqui
A text-to-speech toolkit with many voices and languages, plus voice cloning.
What you need to run it: Python. A GPU for training your own voice; not needed just to speak text.
Read the license first: Weak copyleft — changes to their files must be shared, but your surrounding code stays yours. The company shut down, so the project is community-maintained now.
MyShell and MIT
Clones a voice from a short sample and speaks new text in it, across languages.
What you need to run it: Python and a GPU for sensible speed.
Read the license first: The license permits this; the law and basic decency still apply. Never clone a voice without that person's written permission.
The libraries that sit between your code and a model, plus the tools for fine-tuning one on your own material.
Hugging Face
The standard library for loading and running open models in Python.
What you need to run it: Python. Works on CPU for small models.
PyTorch Foundation
The framework nearly every model here is built on.
What you need to run it: Python. GPU support is an extra install step.
LangChain
Glue for chaining model calls, tools and data sources into an application.
What you need to run it: Python or JavaScript.
LlamaIndex
Turns your own documents into something a model can answer questions over.
What you need to run it: Python, plus somewhere to store the processed text.
Hugging Face
A deliberately small agent library — a few hundred lines you can actually read before trusting it.
What you need to run it: Python and a model, local or hosted.
CrewAI
Runs several agents with defined roles that hand work to each other.
What you need to run it: Python. Costs add up quickly if you point it at a paid API rather than a local model.
BerriAI
One consistent way to call a hundred different model providers, so swapping model means changing one line.
What you need to run it: Python, or run it as a proxy server.
Read the license first: Everything outside the enterprise/ directory is MIT. That directory has its own commercial license, so check which features you are relying on before building a product around it.
hiyouga
Fine-tunes open models on your own examples through a web interface instead of training scripts.
What you need to run it: A GPU. You can rent one by the hour; a consumer card handles smaller models with efficient methods.
Read the license first: The tool is Apache-2.0, but the model you fine-tune keeps its own license — fine-tuning Llama does not free you from Meta's terms.
Unsloth AI
Makes fine-tuning and local running dramatically faster and lighter on memory.
What you need to run it: A GPU, though a modest one goes further than you would expect.
Hugging Face
Fine-tunes a large model by training a small add-on instead of the whole thing — the reason this is affordable at all.
What you need to run it: Python and a GPU.
vLLM project
Serves an open model to many users at once, efficiently. What you use when a hobby project becomes a service.
What you need to run it: A GPU server. Not a laptop tool.
Hugging Face
Wraps a model in a usable web interface in a few lines, so you can hand it to someone non-technical.
What you need to run it: Python.
The unglamorous half: getting your own material into a shape a model can work with, and searching it.
Chroma
A simple searchable store for text meaning, so a model can find the right passage of your documents.
What you need to run it: Python, and it will run inside your app with no separate server.
Qdrant
A fast, production-grade search engine for the same job, when the collection gets large.
What you need to run it: Docker or a binary. Self-host it, or use their free hosted tier.
Meta AI Research
The underlying similarity-search library, when you would rather not run a database at all.
What you need to run it: Python or C++. A GPU build is available.
Weaviate
A search database built for AI, with hybrid keyword-and-meaning search out of the box.
What you need to run it: Docker or Kubernetes.
Read the license first: The license file says plainly that different directories carry different terms. Most of it is BSD-3; some is under their own license. Check the directory your feature lives in.
Microsoft
Converts PDFs, Word files, spreadsheets and slides into clean text a model can read.
What you need to run it: Python, one command. Nothing else.
Tesseract community (originally HP and Google)
Reads text out of images and scans, in over a hundred languages.
What you need to run it: An install, then one command per file.
Explosion
Pulls names, organizations, dates and structure out of text without needing a large model at all.
What you need to run it: Python. Runs fine on a CPU and is very fast.
A license is not legal advice, and I am not a lawyer. If a project is going into paid client work, read the linked license yourself — particularly for anything marked free, with strings.
