Recommended Open Models

You can run an AI model on your own computer, for free, with nothing leaving the machine. No account, no subscription, no company reading your prompts.

We have two practical decisions to take:

  1. Which model
  2. What to run it with

Start Here

If you have never done this before, this is the whole process:

  1. Install LM Studio. It is a normal desktop app for Mac, Windows, and Linux.
  2. Work out how much memory your computer has (see below).
  3. Search in the app for one of the models in our table, download it, and start typing.

That is it. No terminal, no code. If you are not sure where to begin, download OLMo 3.1 7B and see how you get on. It runs on 8 GB.

The rest of this page explains why we picked those models, and what to do when you outgrow the simple option.


Will It Run on My Computer?

Memory is what decides. Not processor speed, not disk space.

Where to look for your number:

Memory What you can run What it feels like
4 GB Small models (3B) Quick questions, short text, phones
8 GB Mid-size models (7-9B) A capable everyday assistant
16 GB The same, with longer documents Comfortable, no compromises
24 GB or more Large models (27-32B) Close to a commercial service

8 GB is enough to do real work. That is the important number on this page. A 7B model compressed to Q4 is about 4.5 GB, so it fits on a modest laptop, an older graphics card, or a second-hand machine. You do not need expensive hardware to get out from under a subscription.

Two bits of jargon you cannot avoid:


What "Open Model" Means

Most models called open are only open-weight. You can download and run them, but the training data and code stay private.

A fully open model releases three things:

Very few capable models release all three. Calling an open-weight release open source is common enough that researchers have a name for it: open-washing.

Openness is a gradient rather than a yes/no. The Model Openness Framework grades it. The OSI's Open Source AI Definition is another reference point, though its handling of training data is still contested.

The same applies to the tooling further down this page. "Open" in a project's name is not a license.


Choosing a Model (August 2026)

Smallest first, because the small ones are good now.

Memory Model Openness Good for
4 GB SmolLM3-3B Fully open Old laptops, no graphics card, phones
8 GB OLMo 3.1 7B Fully open Start here. Everyday use
8 GB Comma v0.1 Fully open, openly licensed data Proving a model can be trained ethically
8-16 GB Qwen3.5-9B Open-weight Coding on modest hardware
24 GB OLMo 3.1 32B Fully open The most capable fully open option
24 GB Qwen3.6-27B Open-weight Coding, data tasks, reading images

The download is a one-off: 2 GB for the smallest, around 4.5 GB for a 7B. After that it runs offline, forever, with no further data cost. On a metered or intermittent connection that matters more than the memory figure.

What sits behind those picks:

Do not assume you need the biggest model you can fit. A 7B is quick and perfectly capable of summarising, drafting, translating, and answering questions. Step up only when the work genuinely needs reasoning. Published benchmarks say little about your own tasks, so try two on real work before settling.

More detail on Qwen3.6-27B

From Alibaba Cloud, under Apache 2.0. It reads images as well as text, so you can feed it a flowchart and ask questions about it.

Qwen reports it beating the previous open-weight flagship, Qwen3.5-397B-A17B - roughly fifteen times larger - on every major coding benchmark, including SWE-bench Verified (77.2 vs 76.2). In practice: near-flagship coding capability on a single domestic GPU rather than a datacenter.

Budget 24 GB. At Q4 the weights alone are around 16.5 GB, so a 16 GB card only manages it with heavier compression or by offloading part of the work to the processor.

It holds about 262,000 tokens at once, its "context window". A token is roughly three-quarters of a word, so that is a long book's worth of text, but not unlimited: a 2 MB text file is about double that and will not fit in one pass.

For evaluating models on your own tasks rather than on published benchmarks, OpenCompass runs reproducible comparisons. For image generation rather than text, PD12M is a public domain image dataset, giving the same provenance guarantee the Common Pile gives for text.


What to Run It With

Running the model yourself is what delivers the privacy and independence. No prompts leaving your machine, no per-token billing, no service to be deprecated.

Most people: LM Studio

LM Studio is a desktop app with model search, download, and chat in one window. Free for personal and internal business use.

The trade-off, worth knowing before you commit: the app itself is proprietary and cannot be audited or modified, even though its command-line tool and SDKs are MIT. It is a convenience layer over open models, not an open tool. If that is unacceptable for your organization, use llama.cpp's built-in interface instead.

Developers and servers: llama.cpp

llama.cpp (MIT) is the engine nearly everything else is built on, LM Studio included. Give llama-server a model file and it serves an OpenAI-compatible API on port 8080 plus a built-in web chat page. No Python environment, no container. Running it directly avoids the overhead a wrapper adds to every request.

Coding agents

OpenCode points a terminal coding agent at any OpenAI-compatible endpoint, including a local llama-server. This is what makes a local model useful for real development rather than only for chat.

Phones

A whole team

Open WebUI is an excellent chat interface. It connects to a llama.cpp backend, serves it over a web page, and installs like an app on mobile. Two caveats: you need a server and a domain, and since v0.6.6 its license adds branding restrictions that make it no longer OSI-approved open source (v0.6.5 and earlier remain BSD-3-Clause). Fine for internal use, but check the terms before redistributing or rebranding.

A note on Ollama

Ollama is the usual recommendation for an easy start, and we no longer suggest it as a first choice. It has a history of crediting llama.cpp poorly despite being built on it, and now measurably trails llama.cpp on throughput by anywhere from 10% upwards. LM Studio covers ease of use and llama.cpp is straightforward to run directly, so the wrapper has little left to offer.


A starting point. We will expand this page as we test more of these ourselves.