Models
Search across registered AI nodes and imported catalogs — pick a model on the left, its full card is on the right. Claimed values are never blended with measured ones.
Models you download and run yourself with Ollama: no server, no account, and your data stays on your machine. There is nothing to verify remotely, so they are ranked by popularity (Ollama pulls); memory needs are 4-bit estimates.
Ask the advisor — describe a task, get the best node + model, priced and explained
Select a model from the list.
Offerings
Playground runs in your browser — your prompt goes straight to the node, never through this registry
Rankings
Rolling registry measurements of live endpoints — these rank speed and reliability, not model intelligence. Nodes are blind-probed at unpredictable times; imported catalog entries are unprobed and excluded. Adoption and Community rankings arrive with settlement and reviews.
Nodes
The supply side, first-class: every provider serving through OpenNodes, with the trust tier it has earned. Verified means identity-fingerprinted, context-verified, capacity-measured, and continuously probed — the qualification procedure is automated and the same for everyone. Region lists the countries where a provider has known sites, not where each of its models runs: the map has that detail.
Where models run
Every known serving site by country, with the models it serves and how far each location can be trusted: claimed (the provider or a news report says so), documented (the provider's own dated page), measured (registry probes), attested (a signed card backed by an audit). Built from node cards, imported catalogs and curated seeds; corrections welcome. Also as JSON, CSV and GeoJSON for QGIS, ArcGIS or any web map. Each site also shows the provider's official residency commitment where it covers that site: processing (inference and stored data stay in the area), storage (data at rest only), or none; every provider's commitments are listed below the table.
Legend what the markers, chips and columns mean
MARKER COLOUR · EVIDENCE
- measured or attested: registry probes or a signed, audited card
- documented: the provider's own dated page
- claimed: the provider or a news report says so
MARKER SHAPE
- hollow: approximate point, only the region or country is known
- dashed: planned, not serving yet
- Size grows with the models served. Several sites in one city share a marker; click it for each one.
RESIDENCY · WHAT THE PROVIDER PROMISES
- PROCESSING prompts, outputs and stored data stay in the area
- STORAGE data at rest stays; processing may not
- NO FIXED AREA the provider says data may go anywhere
- NOT STATED no public promise found
- opt in: a setting or regional endpoint you choose · enterprise: contract or paid plan only
COUNTS
- Providers: companies with at least one site in the country
- Models listed: distinct models a provider says it serves there. A dash means the provider names its sites but not which models run at each
- Click a country to filter the map and the site table.
Countries
Residency commitments
What each provider officially promises about where prompts, outputs and stored data stay, with the page that says so, including routers and providers with no known sites. default: every customer; opt in: a setting, regional endpoint or region you choose; enterprise: contract or paid plan only. Curated from public pages; a promise is not proof (ONP-2 §3b). Also as JSON and CSV.
New
Dated listing events, newest first — registrations, tier changes, suspensions — each with its provenance. A node that keeps dropping offline and coming back shows as one availability line. Also available as JSON for agents (every transition).
Chat
Straight from your browser to the nodes you pick — your messages never pass through this registry. Add several models and switch modes: compare them in parallel, let a jury judge their answers, or chain them into a pipeline.
Add a model on the right and say something.
Model tests smoke probes for the current roster — quick sanity checks, not benchmarks
scenario instructions
scenarios — save & load
Saved in your browser — no account. Export/import use the portable ONP-6 §5 format; a share link carries the whole scenario in the URL (never your API keys).
Surge, a desktop client with native OpenNodes support (models + MCP tools in one app), is in pre-release: download the Windows installer, or get the source on GitHub.
Available today: bring every model here into Cursor, VS Code, or any OpenAI-compatible app via the onp gateway and MCP server.
Build an AI node
From "which model?" to a registered node: pick a model and quantization, choose a runtime and hardware, and get a ready Node Kit configuration. Estimates are rough sizing guidance, not measurements — the qualification battery measures reality after you register.
1 · Model & quantization
2 · Runtime
Simplest path — one install, GGUF quantization built in. Best for personal nodes and development; modest peak throughput.
Production serving — continuous batching, paged attention, best multi-user throughput. The default for a serious node.
NVIDIA-optimized kernels — lowest latency on H100/H200 class hardware; heavier build and engine-compilation step.
3 · Hardware
Where to run it: your own hardware, a university/institutional cluster, or GPU clouds (Google Cloud, AWS, Hetzner, OVHcloud, CoreWeave, and regional providers). Data-residency requirements narrow the list — that constraint is a first-class filter in this registry for a reason.
4 · Your node
select a model and runtime to generate your configuration
Then: install the Node Kit → onp-node init with this config → onp-node check → register. Trust tiers are earned through the automated qualification battery.
Register an AI node
Anyone can register — a lab, a university cluster, a single Ollama box. Registration is open; trust is earned through the automated qualification procedure. Serve a signed Node Card at /.well-known/open-node.json (the Node Kit does this for any vLLM/Ollama/TGI deployment), then submit it here.
The qualification procedure
Listing is open; tiers are earned. The registry runs an automated admission battery against your live endpoints — no paperwork, no gatekeepers, no exceptions. Every verdict is recorded and visible in your node's listing history.
PUT /v0/nodes/{id}/ping; this registry does not yet re-run it on a timer.onp.openai.chat/v1). A tier covers only the offerings it could test: speech, embeddings, image or any other offering is listed as UNTESTED, and its tier is capped to UNVERIFIED whatever the node's tier. A node with no chat offering stays unverified, because Stage A has nothing to call.Probe load
Everything this registry sends to your endpoints, as the code runs today. A chat offering is one bound to onp.openai.chat/v1; other offerings get only the health and models checks. Probe calls carry your probe key if you gave one at registration, and the registry does not pay for them.
PUT /v0/nodes/{id}/ping. No timed re-run yet.Sends: 1 GET to your health endpoint, 1 GET to
/models, and 1 blind chat completion per chat offering (a short prompt, at most 32 output tokens).POST /v0/nodes/{id}/admit), at most once every 30 minutes per node on registry.opennodes.io.Sends, per chat offering (up to 20 chat completions):
· Fingerprint battery: 3 fixed prompts at temperature 0, at most 64 output tokens each. Skipped when the registry holds no reference outputs for the model.
· Calibration: 1 call with about 4,000 characters of filler, at most 4 output tokens, to measure your tokenizer.
· Context: 3 probes sized at 25, 50 and 90% of the claimed context window, at most 64 output tokens each: about 1.65 times the window in prompt tokens altogether (about 210k tokens for a 128k window).
· Capacity ramp: one batch each of 1, 4 and 8 parallel requests (13 calls, at most 48 output tokens each) inside a 3-second budget; a step that would start after the budget is skipped.
Sends: 1 GET to your health endpoint every 5 minutes on average (each gap is random, 2.5 to 7.5 minutes), and 1 blind chat completion per node per hour on average (30 to 90 minutes apart, to a randomly chosen chat offering, at most 32 output tokens), each after a fetch of your card. That is about 288 health checks and 24 chat calls a day per node, whatever the number of offerings.
Sends: 3 TCP connections to port 443 of your API host from each probe site (Nuremberg and Helsinki). No HTTP request.
Principles
Claimed ≠ measured ≠ attested. Self-reported values are labeled CLAIMED; registry measurements override them and show as ● MEASURED; third-party Verifiable Credentials earn attested. The three are never blended.
Transparent rank, by rule. Every result carries its rank explanation — component weights and values. Undisclosed ranking inputs are non-conformant; sponsored placement, if ever present, must be labeled and excluded from rank.
Probe load is bounded and published. The probe load list above is all the traffic the registry sends you. A card may declare a probe allowance (payment.probe_allowance); this registry does not read it yet, so the load above applies whatever the card says.