AI models and the GPUs that run them

A record of 3,567 machine learning models — who built each one, where, when, how large it is, and whether its weights were published. It is broader than a list of things you can run: a model nobody can download still has a specification worth reading, and knowing what exists is part of judging what is worth running.

What makes this catalogue different from a plain index is that it is attached to a hardware calculator. Every model whose weights are open and whose size is on record has been assessed against every graphics card we hold specifications for. Open one and it tells you which cards can run it, how much memory each needs, and roughly how many tokens per second to expect.

The single most consequential fact about a model is whether its weights were published. A model's weights are the trained values that make it what it is — hundreds of billions of numbers arrived at over months of computation. Publishing them means anyone can download the model and run it on their own hardware, permanently, with no account and no network connection. Not publishing them means the model exists only as a service, reachable through an interface its owner controls.

That distinction decides everything else on this site. A closed model cannot be run locally at any price, so no amount of graphics card changes the answer and no memory figure applies to it. An open-weight model can be run by anyone who has hardware large enough to hold it, which turns "which model should I use" into a question with a hardware answer.

"Open weights" is not the same as open source in the sense the term usually carries. The weights being downloadable says nothing about the training data, the training code, or the licence attached to the result — some published models permit commercial use freely, others restrict it. What is consistent is that the model runs on your machine rather than someone else's.

Of the 3,567 models here, 1,335 have open weights and 2,232 do not — roughly 37 in every hundred is something you could download today. Where the licensing was never recorded we count it as closed, on the principle that an unstated licence is not one to rely on.

Parameters are the learned values inside a model — the numbers adjusted during training that encode everything it knows. A model described as 8B has around eight billion of them, and each one occupies space in memory. That is why the parameter count is the column that matters most here: it is the closest thing to a direct statement of how much hardware a model needs.

The relationship is close to linear, but the constant depends on how the model is stored. At full precision each parameter takes two bytes, so an eight-billion-parameter model needs sixteen gigabytes before anything else is accounted for. Compressed to around four bits — which is what most people running models locally actually use — the same model needs closer to five, at a modest and usually acceptable cost in accuracy.

Two things complicate it. Mixture-of-experts models route each token through only a fraction of themselves, so they generate text at the speed of something far smaller while still needing every parameter resident in memory — fast, but not cheap to hold. And the conversation itself takes memory that grows as the conversation does, which is why a model that fits comfortably for short questions can fail on a long document.

Bigger is also not automatically better, and increasingly is not. Capability per parameter has improved sharply enough that a recent model at eight billion parameters routinely outperforms an older one several times its size, while fitting on hardware the older one never would. Sorting this catalogue by release date is usually a better guide than sorting it by size.

Almost every model in this catalogue descends from a single architectural idea published in 2017, in a paper called "Attention Is All You Need". It introduced the transformer, which processes a whole sequence at once and learns which parts of it bear on which — replacing the step-by-step recurrent designs that came before, and scaling far better than they did. Every model listed here that generates text is a transformer or a close relative of one.

What followed was mostly a demonstration that the architecture kept improving with scale. Successive models grew by orders of magnitude and got correspondingly more capable, which set the pattern for several years: more parameters, more training data, more computation. By 2020 the largest models were capable enough to be commercially serious, and they were reachable only as paid services — the weights stayed with whoever trained them.

The turn came in 2023, when several organisations began publishing weights for models good enough to be worth running. That changed the question from "which service do I subscribe to" into "what can my own hardware do", and it is the reason a site like this one exists. Before it, the hardware question had no meaningful answer for most people.

The years since have been less about scale and more about efficiency. Better training methods, better data and architectural work such as mixture-of-experts routing have all pushed capability per parameter upward, while quantisation — storing each weight in four or five bits instead of sixteen — cut the memory a given model needs by roughly three quarters. Between them, models that would once have required a datacentre now run on a desktop graphics card, which is the practical development this whole catalogue rests on.

The largest model on file is BaGuaLu at 173.9T parameters. At that scale a model is datacentre infrastructure rather than something a workstation holds, and the figure is more useful as a marker of where the frontier sits than as a shopping decision.

The catalogue covers work from 1,314 organisations across 50 countries. That spread is worth noticing on its own: model development stopped being concentrated in a handful of laboratories some time ago, and several of the most widely used open-weight models come from organisations that did not exist a few years earlier.

Parameter counts are recorded for 2,315 of the 3,567 entries. The gaps are mostly closed models whose owners never published a size, and occasionally open ones nobody measured. A model without that figure is listed but cannot be given a memory or speed estimate, since every one of those calculations starts from it.

Entries span 1950 to 2026, so the catalogue is a record of the field rather than a snapshot of what is current. Filtering by release year is the quickest way to narrow it to models worth actually considering.