About OpenSourcesAI
Helping you choose and run local AI.
OpenSourcesAI helps you work out what runs on your hardware, choose the tools you need, and build a useful local AI setup. It's built and run by CJ Zwart, with help from AI.
Who runs OpenSourcesAI
Editor and founder, OpenSourcesAI · Sebastopol, California
CJ Zwart runs OpenSourcesAI from Sebastopol, California. It is a one-person operation, built with help from AI. The site began as a directory of local AI tools and models and has become a set of tools, estimates and measurements that answer one question: what actually runs on the hardware you own.
Where this started: CJ's Home Assistant story, the smart home project that led him to open source and to using local AI for code.
The site combines estimates from published model files and hardware specifications with measurements from CJ's RTX 3080 Linux rig and RTX 4070 Ti Windows desktop. His M1 MacBook Air is used for Apple memory measurements and for detection and compatibility checks. Published measurements link to their result files, which record the runtime version and the exact model file used.
CJ also builds and maintains @opensourcesai/bench, the open-source command-line tool the site uses to measure tokens per second, VRAM residency and usable context on a local machine; @opensourcesai/cmdcenter, a free, open-source local dashboard for your Ollama machine; and the fit engine behind the Local LLM Compatibility Checker.
Commercial relationships are disclosed on the pages where they appear and never change a verdict. Corrections are welcome and are published with a dated note.
GitHub · @opensourcesai/bench · @opensourcesai/cmdcenter · LinkedIn
What the site gives you
Tools and reference pages for each step, from finding out what your hardware can run to measuring what it actually does:
- Compatibility Checker: See which models fit your GPU, VRAM and RAM.
- Model Wizard: Answer a few questions, get a model and stack direction.
- AI PC Builder: Pick a model size, workflow and budget, then get the parts to buy.
- Models: Open-weight model profiles with hardware fit and copy-paste run commands.
- Hardware: GPUs and machines rated for local AI, with honest VRAM tiers.
- Tools directory: Reviewed AI tools, filterable by workflow, license and deployment.
- Benchmark: Measure what your running setup really does. Nothing leaves your machine.
- Command Center: A local dashboard for your Ollama machine: real hardware, real VRAM residency.
- Start Here: The beginner path: what local AI is and how to begin.
The iPhone app is in free public beta, with no account or sign-up. Download and test it through Apple's TestFlight.
How the information is produced
The site works with three kinds of information and keeps them apart.
Estimates
The Compatibility Checker, the PC Builder and the hardware tiers work out fit from published model file sizes and hardware specifications. They are estimates, not measurements.
Sourced information
Model, tool and license details come from primary sources such as model cards, official documentation and project repositories.
Measurements
Measured figures, such as tokens per second, VRAM used and usable context, come only from runs on CJ's own machines. Each one links to its result file, which records the runtime version and the exact model file used.
- 2570server: RTX 3080 10 GB, Linux. The dedicated GPU test rig for runtimes, models and validation runs.
- Windows desktop: RTX 4070 Ti 12 GB, Windows 11. The second measurement point for checker and bench runs.
- MacBook Air M1: Apple Silicon unified memory. The macOS and Apple Silicon detection target.
Every quarter the site re-checks model availability and licenses, Ollama tags, tool projects, prices and partner websites, and publishes the results, including what is still open, in the latest re-audit.
More detail: How this site is produced and the methodology.
Independence and corrections
No tool vendor, model developer or cloud provider approves what the site says. Partner and affiliate links are labelled on the pages where they appear, and a commercial relationship never changes a verdict, a ranking or a recommendation. The build fails if a monetized tool has no partner record with its disclosure.
Corrections are welcome. Send them through the contact page, with an official source link for anything about hardware specs, licenses or pricing. Material corrections are published with a dated note on the page.
The full policies are in the editorial standards. For sponsorship and media, see Advertise with OpenSourcesAI.