Skip to content
Weights Not Walls
Menu

About

Who runs Weights Not Walls, who pays for it, who it works with, and where its data comes from.

Weights Not Walls is a site about open-weight AI models: models whose weights you can download and run yourself. For each model it tries to answer five questions in one place. Which model suits this job? Will it run on my machine, and in which format? How good is it, according to whom? Where can I run it without my own hardware, and at what price? What came out this week?

What exists today

  • A catalogue built from the Hugging Face Hub. The snapshot of 19 November 2026 indexes 4,715 models, of which 390 are reviewed (the rule is below).
  • A page per model with its size, licence in plain words, local formats and whether they fit common machines, API providers and their prices, and scores with their source.
  • A page per evaluation set, and a page on how we read scores.
  • This blog and a weekly digest, also as an RSS feed.

What does not exist yet

  • Our own evaluation runs. Every independent score on the site today was produced by someone else, and the page says who.
  • Trying a model in chat from its page, one API key for every open model, fine-tuning, creator payouts and the fund. The model pages mark these as coming.
  • The newsletter sign-up, the community programmes and the advisory group of model makers.
  • Automatic refreshes. The data is a snapshot, refreshed by hand for now, and it can be a day or more behind the Hub.

Who runs it and who pays

The site is run by the Solana Foundation and funded by Solana AI (the Solana Foundation). The Foundation takes nothing from it. Any fee the site ever collects goes to a public fund for new open models. The site collects no fees today.

There is no token. You do not need a wallet to use the site, and today you do not need an account either. Solana appears only where it does real work: the fund's payments will be made and published on Solana, so anyone can check them.

A funder is a conflict of interest. The independence rules say how we handle it.

Partners

Status means: in place, working today; in discussion, we are talking and nothing is signed; planned, we intend to ask and nothing is agreed. Where a partner's service appears on a page, the page will say so.

Who What they do or will do Status
BlockRun Routing and payments: one API key for open models, paid by card or per request. In discussion
Unsloth Fine-tuning, and local versions of new models on the day they come out. Planned
pay.sh Payments for software agents that call models. pay.sh is run by the Solana Foundation. Planned
Nito and Tinfoil Private chat. Solana AI's confidential chat at solana.ai/chat runs models privately with Nito and Tinfoil. The chat exists; links to it from model pages are planned. Chat in place, links planned
Akash Compute, for fine-tuning, for bounty work and for people who want to contribute machines. Planned

Hugging Face and Epoch AI are data sources, not partners. We have no agreement with either. We use their public data under its terms and credit them on every page that shows it.

How a model becomes "reviewed"

Everything on the Hub in the text-generation and image-text-to-text categories can have a page. Most are marked unreviewed: shown, searchable, and limited to what the Hub data supports. A model is reviewed when all of these hold:

  1. It is the maker's own repo, from an organisation on our list of original makers. The list has 72 organisations, each checked against its page on the Hub.
  2. It is a base model or the maker's own instruct or chat model, not a quantised conversion, an adapter or a merge.
  3. Its licence is one we have read and classed (permissive, open with conditions, non-commercial, or custom).
  4. The Hub gives its parameter count.
  5. It is tagged text-generation or image-text-to-text, and the repo is not disabled.
  6. People use it: at least 100,000 downloads or 500 likes, or at least 20 likes if it is less than 21 days old, or it is served by an API provider.
  7. We have not excluded it by hand. We record every hand exclusion or inclusion with its reason.

The rule is applied by a script. Before launch a person will go over the whole reviewed list; until then, "reviewed" means checked against our published rule, not checked by hand.

Where the data comes from

  • Model data: the Hugging Face Hub API, read anonymously. Each model page links to its repo.
  • Providers and prices: Hugging Face Inference Providers, read from the router's model list, with the time the data was fetched.
  • Independent scores: Epoch AI's benchmark data, under CC-BY 4.0. Rows that Epoch AI ran itself are labelled "Re-run by Epoch AI"; rows Epoch AI collected from other leaderboards are labelled "reported" and link to the leaderboard. Citation: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource].
  • Claimed scores: the maker's model card, where it lists scores in a form a program can read.
  • Artificial Analysis and Arena: we link to them from model pages. We do not copy their data.

Sizes, memory needs and "will it run" verdicts are worked out by us from the files on the Hub. The method, with its formula and its limits, is on the method page.

Contact

To report a mistake or a problem with a listing, see the content policy. Contact: hello@weightsnotwalls.com.