Open weights: list where the model can be served
Published weights mean many sellers. The page buyers want is a dated list of hosts, the build each one serves, and the limits each one enforces on the model.
The question after the download link
A model with published weights is a model with many sellers, and the next question a buyer asks is where to run it without running it themselves. The maker is the best placed to answer: keep a dated list on the model page of the hosts that serve it, which build each serves (precision, context limit), and a link to each host's own page for the price. The weights link answers the engineer. The host list answers the person who signs the invoice.
A maker that stops at the weights leaves the buyer to search, find hosts that may be serving an older version or a smaller window, and guess which is which.
The list does not have to be long to be useful. A short one, dated and accurate, answers the question a buyer was about to put to a search engine, from the source best placed to answer it.
What the market already shows
Our token price table holds 93 models sold by more than one seller. We cannot tell from that data which of them have published weights, and do not claim to. What it does show is that, for a model with several sellers, the same name can carry different prices, different limits and sometimes different builds.
The inference layer lists 186 live products, from hosted inference APIs to the serving engines buyers run themselves. A buyer choosing among them for one model needs the maker to say which of them actually serve it.
The price table puts each seller's rate for a model in one row where the names match, and a buyer reading it wants to know whether the cheapest seller is serving the same build as the dearest. Only a source closer to the model can answer that.
The columns a host list needs
Host name and a link to the page where that host lists the model. The model identifier the host uses, which may differ from yours. The precision served: 16-bit, 8-bit, 4-bit or a named quantisation. The context window the host enforces, which can be shorter than the model supports. Whether the host serves your latest version, and the date you last checked.
Leave the price out, or link it rather than copying it. A copied price goes out of date the first time the host changes it, and makes your page look careless. We hide any price older than 21 days for the same reason, and a link to the host's own dated page stays correct for as long as the host keeps it current.
Leave rankings out too. A list ordered by your preference reads as an advert. Alphabetical order and the same columns for every row read as a reference.
Mark hosts that serve a fine-tuned or modified version under your model's name. The name is read as meaning your weights, and a row that is something else needs to say so.
Self-hosting belongs on the same page
Some buyers will run the weights on rented GPUs. For them the list continues with the serving engines you have tested, the version of each, and the memory the weights need at each precision. That last figure is arithmetic from the parameter count, and the GPU rate table is where they price the card it implies.
Say which engine features you tested and which you did not. A model that works in one engine with a particular chat template and fails in another produces complaints that reach the maker, and a sentence saying which combinations you checked heads many of them off.
Say that the key-value cache needs memory on top of the weights and grows with context and batch. A buyer renting an 80 GB H100 SXM (lowest listed hourly rate: $1.68) needs to know before starting whether the build they chose leaves room for the work.
What the list does for the maker
A buyer who cannot find a host for an open weights model does not have to rent a GPU instead; a model already offered by a provider they use is the easier choice. The host list is how a maker reaches buyers who will never download a file, and for whom a per-token price may be the only form that fits.
It also gives the maker a place to record what hosts get wrong. If a host serves a shorter window or an older version, a dated list that says so is a quieter correction than a public complaint, and buyers can see the difference for themselves.
Keeping the list current without a team
A host list goes stale the same way a price does. Put a checked date at the top and recheck on a schedule you can keep, monthly or at each release. Remove hosts that stop serving the model rather than leaving dead links. Ask hosts to tell you when they change the build they serve.
If a host asks to be added, ask for the page where it lists the model and the build it serves, and add it once that page exists. A list built from pages anyone can check can be verified by the buyer; one built from emails cannot.
None of this changes which model is better. It shortens the time between a buyer deciding to try your model and their first request. Naming matters here as well: a list is only useful if every host writes the identifier the way your naming page says.
What we could not answer
- ⚑This site's data cannot say which models have published weights, so the count of models with several sellers is not a count of open weights models.
- ⚑We do not measure whether hosts serve the same build, so the advice on checking builds rests on what hosts state.
What this rests on
infratoapp, “Open weights: list where the model can be served”, read 27 October 2026. models sold by more than one seller: 93. https://www.infratoapp.com/blog/open-weights-list-where-the-model-runs
Questions people ask about this
Where can I run an open weights model without my own GPUs?
On a hosted inference provider that serves that model and bills per token. Check the maker's page for a list of hosts, then each host's own page for the price, the precision served and the context limit it enforces, which can differ from the model's own maximum.
Why do hosts serving the same model give different results?
They may serve different builds. One host may run the weights in 16-bit and another a 4-bit quantisation, or apply a different chat template or a shorter context window. Any of those can change answers. The host's page should state each one.
Should a model maker list host prices on the model page?
Linking is safer than copying. Hosts change prices on their own schedule, so a copied figure goes stale quickly. A link to each host's dated pricing page stays correct while that host keeps its own page current, and avoids the list reading as an endorsement.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
A license name tells a buyer where to look, not what they may do. Put the permissions in a short can and cannot list beside the name, dated, on the model page.
Assistants quote pages that state facts plainly: an answer first, exact names, dated figures and structured data. None of that guarantees a citation.
When sellers spell one model several ways, price tables split it into separate rows. Publish one canonical name and a grammar for its variants, and keep both.