Methodology: how we collect, check and classify open AI models
How ai-opensource.com decides what to list, where facts come from, how memory estimates and license categories work, and how often pages are updated and corrected.
In short. We list only models whose weights anyone can download, take facts from primary sources, estimate memory with a published formula, classify licenses by what the license file says, and show an update date on every page.
What we list
A model is in the catalog if the public can download its weights, including gated downloads that ask you to accept a license first. We do not list API-only models, leaked weights or announcements without weights. The catalog is curated rather than exhaustive: we focus on models that are current, widely used or technically notable, and we mark older ones as legacy instead of deleting them.
Where facts come from
Every model page cites its sources, and so does every news item. We use primary sources in this order:
- The license file in the model repository.
- The model card, the configuration files and the real file sizes of the checkpoints.
- The developer’s announcement or technical report.
- Platform metadata such as the Hugging Face API, and the project’s code repository.
We do not use rumors, social-media posts or aggregator pages as the only source for a number. When sources disagree, we say so on the page. The release date is the day the developer announced the model or published the weights, not the day a repository was created.
Benchmarks
We do not run our own benchmark suite, and we do not rank models by a single score. Benchmark numbers on this site are reported by the developers, with the test setup they chose, and are labeled that way. How to evaluate AI models yourself explains how to test a model on your own work.
How we estimate memory
The estimates are deliberately simple and conservative. They are for planning, not guarantees.
| Format | Bytes per parameter |
|---|---|
| FP32 | 4 |
| BF16, FP16 | 2 |
| FP8, INT8 | 1 |
| INT4, FP4 | 0.5 |
| MXFP4 | 0.53 |
| NVFP4 | 0.56 |
| Ternary (weights of −1, 0 or +1, about 1.75 bits per weight) | 0.22 |
| GGUF Q4 (Q4_K_M, about 4.8 bits per weight) | 0.6 |
Memory for the weights is parameters multiplied by bytes per parameter. We add 15% for runtime buffers and a modest context cache. Then these rules apply:
- Mixture-of-experts models keep every expert in memory, so we use total parameters, not active ones. Active parameters only affect speed.
- Language models are listed with a 4-bit estimate, because they are routinely quantized. If the developer ships 4-bit or ternary weights, we use the actual download size plus 10% instead.
- Image, video, speech, music and embedding models are rarely quantized, so we use the weights as shipped plus 15%. Pipelines with several components, such as a separate text encoder, use the total download.
- A developer’s own minimum overrides our formula whenever the developer states or measures one.
- Long contexts need more. The key-value cache grows with every token, so a long prompt can need far more than these numbers. The memory guide has a worked example.
We then place each model in a hardware tier by that number:
| Tier | Up to |
|---|---|
| Laptop or phone | 16 GB |
| Gaming GPU | 32 GB |
| Workstation | 96 GB |
| Big-memory desktop | 512 GB |
| GPU server | 1,536 GB |
| Multi-node cluster | More than 1,536 GB |
How we classify licenses
We read the license file itself, not only the label on the model page, and we note when the two disagree. Each license falls into one of three types:
- Permissive: you may use, modify and sell the model, and the conditions are light, such as keeping the notice. Apache 2.0 and MIT are examples.
- Conditional: commercial use is allowed, with extra conditions. We tag them: attribution, a hosting agreement, a revenue cap, territory limits, a usage policy, a security review, output filtering or naming rules.
- Non-commercial: research and personal use only. Commercial use needs a separate agreement.
We also show whether the Open Source Initiative has approved the license, and the date we last verified the text. This is a reading aid, not legal advice. AI licenses explained gives you a checklist to use before you ship.
Lists, comparisons and the featured set
Best-of lists and comparisons are editorial choices. Each one states its criteria and gives a reason for every pick. The featured models on the home page are a curated set of about twenty, chosen to cover a range of sizes and categories. Being featured is not a ranking. When the catalog changes, we review the lists that mention the affected models.
Updates and corrections
Each page shows the date its facts were last checked. News items show the date of the event and the date we last verified them. When we find an error we correct the page and update its date, and if a correction changes a conclusion we say so on the page. Please report mistakes through the contact link in the footer.
Open data
The catalog is published as open data: models in JSON and CSV, and licenses in JSON. Text and data are released under CC BY 4.0, so you can reuse them with credit.
Languages and AI assistance
English is the reference version, and translations follow it. We use AI tools to help with research and drafting, but we do not publish a number we cannot trace to a source. Every model page carries at least one source link, and news items list theirs.