Method

What is listed

The list covers open-weight models whose weights can be downloaded from the publisher's own Hugging Face organization. Each model family appears once, at its latest version. When a lab ships a new version, the older one leaves the list and its address redirects to the new page. Sizes of the same generation share one page. A preview release is listed and marked Preview. A model that has only been announced is not listed.

Dates are the publisher's release date when one is given, otherwise the day the repository was created; each page says which. Parameter counts are the total in the weight files of the listed repository. For a mixture-of-experts model that is every expert, not the active part. Licenses are the identifiers from the repository. Read the license text before you build on a model.

The list was last checked on 2026-10-11: 154 models from 44 labs.

Weight options

A weight option is one size, one format and one quantization level, for example 8B, GGUF, Q4_K_M. Options come from file listings on Hugging Face, collected on 2026-10-11: 503 in total. This round lists only files published by the lab itself. Community conversions are not listed yet.

For original weights the size is the total of the weight files in the repository, which can include more than one variant. In GGUF repositories, projector, calibration and auxiliary-head files are not counted as options.

Minimum memory

Memory need belongs to a weight option, not to a model name. A figure is marked ● when the publisher states it and ◇ when it is our estimate. Everything else says Unknown.

Estimates exist only for text language models: the Language models, Coding, Agents and Translation tasks. The estimate is the weight file size, plus the key-value cache for a 8K-token context, plus 0.5 GB for the runtime. The cache per token is 2 × layers × key-value heads × head dimension × 2 bytes, read from the repository's config. When the config does not give those numbers the cache is left out, and the note on the option says so. A longer context needs more memory than the estimate.

Speech, image, video, music and OCR models get no estimate. Their memory use depends on resolution, length and pipeline, and file size is a poor guide. 64 of 154 models have a figure; 90 say Unknown.

Fits, tight, too large

Set your device memory and each option is compared against it. Fits: it needs at most 85% of your memory. Tight: more than 85%, up to 100%. Too large: more than 100%. Unknown: there is no memory figure. This is a memory check only and says nothing about speed.

For a graphics card, enter its video memory. For Apple silicon, enter the unified memory. For CPU only, enter the RAM. MLX files are counted only for Apple silicon. The page detects the device type and shows the graphics chip name as a hint. It never guesses memory. Your settings stay in your browser.

Third-party scores

Some pages quote scores from other evaluators, with the source and the data date each time. They are not this site's measurements. They were run on hosted endpoints, so they do not describe a particular quantized file. Scores from different sources are never put in one column or one chart. 19 models have at least one.

Charts

A chart plots one metric from one source against minimum memory on a log scale. It appears only when at least 8 models have both a score and a memory figure within the chosen memory budget. Otherwise the page shows the table alone. Because the scores are from hosted endpoints, each model is one point, at the memory need of its original weights. The line connects points that no other point beats on both memory and score. It does not predict anything in between.

Tests on this site

This site has not run any model yet, and every model page says so. Current status: Not tested yet 130, Not in the first test round 24.

The first round will use hosted APIs where they exist and rented servers otherwise. Each result will name the endpoint or the hardware. A speed measured there is not a speed on your machine and will not be shown as one.

Tasks that generate something are run 3 times and every run is shown. Tasks with a checkable answer are run 10 times. The first round does not use a language model as the judge. Code, agent, embedding and safety models are not in the first round; their pages list sources only.