Read the machine details
Compare publisher-declared accelerator, VRAM, price, context, and performance conditions.
Public browsing
Continue with your RMA account.
OPEN MODEL ENDPOINT LAB · COMMUNITY HOSTS
Open models. Real endpoints. Honest machine details.
Discover independent model hosts, compare publisher-reported hardware and performance, and see optional activity from real requests.
Listings are free. RMA does not check, verify, probe, or test endpoints. Hardware, performance, model, and capability details come from publishers. RMA activity appears only from successful requests people chose to route through RMA and share.
Know where signals come from
Publisher claims, opted-in RMA-routed activity, and community reports stay clearly labeled.
Choose your destination
Review the host, model, data practices, price, and route limits before creating a local client connection.
Explore model endpoints
Filter the directory, then add up to three hosts to a private, in-browser comparison.
FREE COMMUNITY LISTINGS
No public model endpoints yet
Be the first to list an open-weight inference endpoint. Add your model, hardware, limits, price, and data practices. Publishing is free; RMA does not test endpoints.
Start a private provider draft ↗FOR MODEL HOSTS
Put your endpoint in the open.
Give people enough information to choose your service with their eyes open.
Make the offer legible
Create a provider profile and add a public HTTPS API endpoint. Declare the exact served model ID, upstream release, hardware, runtime, capacity, limits, price, and data practices. Include dated measurement details when sharing performance. Individuals and organizations are welcome.
List for free
Publishing is free. RMA does not check, verify, probe, or test endpoints or publisher claims. No ownership challenge, synthetic request, compatibility test, scheduled monitor, or RMA benchmark is run. Review declarations and community reports yourself before connecting.
A listing does not provide compute, provider credits, or inference. Hosts supply and pay for their own hardware, API access, and provider costs. People who connect choose what to send and remain subject to the host and provider terms.
Show activity by choice
When someone creates a connection, they may separately opt in to activity samples linked privately to an opted-in connection from successful requests they choose to route through RMA. RMA stores a private connection reference, endpoint, offering, and sample time in a 30-day rolling window; public counts appear only after five samples across at least three UTC days. Stored samples can be linked to your connection inside RMA, but public totals contain no account or connection ID. Prompts, responses, and account IDs are not included. This is not an uptime check or reliability score. Sharing can be stopped from the connection dashboard.
Update your declarations or pause a listing whenever needed. Respond to community reports from your dashboard.
HOW TO READ A LISTING
Every signal has a source.
Hardware, model identity, capabilities, price, performance, availability, and data practices are publisher declarations. RMA activity counts only successful user-initiated routes from connections that opted in; it does not measure unique people, uptime, comparative speed, answer quality, or safety. Reviews and votes are community reports.
Direct coding-client routing is free within service limits. RMA charges no per-call service fee for optional remote MCP model help. Inference may incur charges from the host’s provider. RMA does not pay endpoint hosting costs or offer paid placement.