Documentation · Reference
Glossary
The marketplace's vocabulary, one line each. Terms link to their home page.
| Term | Meaning |
|---|---|
| Account | The money: a credit balance, a ledger and an account-wide spend limit. Personal or organization. |
| Author | The creator of a model's weights — always who made them, never who serves them. The first half of a model id. |
| Capacity | A provider's live report of concurrent request slots (current of total). Fresh reports steer routing; absence is neutral. |
| Catalogue | The public model list and everything declared or measured about it. GET /models and friends. |
| Completions | POST /completions — plain text in, text out, no messages. The older sibling of chat completions. |
| Context length | The maximum number of tokens an endpoint accepts, prompt plus completion. Per endpoint, not per model. |
| Credit hold | The conservative estimate reserved before a request runs; settles to the actual cost when it finishes. |
| Credits | The prepaid balance. $1 of stablecoin becomes $1 of credits; requests draw it down. |
| Data policy | A provider's declared training and retention_days for an endpoint. Filterable with data_collection: "deny". |
| Declared | Stated by a provider about its own endpoints (prices, context, policies). Never taken as measurement. |
| Deposit | A stablecoin transfer that credits an account balance. Matched by wallet binding or reported to support. |
| Dialect | The API shape a source speaks: OpenAI-compatible or Anthropic-compatible. Chosen per source. |
| Endpoint | One provider's serving of one model: its own price, context, quantization and measured performance. The unit routing chooses between. |
| Family | A base model and its variants, grouped for browsing. Requests name models, never families. |
| Fallback models | The models array — up to 20 ids tried in order when the primary model cannot be served. |
| Generation | One inference request, identified by its X-Generation-Id and carrying a billing record. |
| Health | An endpoint's tier from measured uptime: healthy, degraded, down, unknown. Routing prefers healthier tiers. |
| Ledger | The account's money history: deposit, usage, refund, adjustment rows, exact decimals, newest first. |
| Market share | An author's fraction of all tokens processed in the rankings window. |
| Measured | Computed by Heimdallm from real traffic and probes: uptime, TTFT, response, throughput, tool-call success, benchmarks. |
| Model id | author/slug, lowercase. What a request passes as model; matched case-insensitively. |
| Playground | The browser chat interface. Calls the API on a key you hold; history stays in the browser. |
| Provider | An independent operator serving models for payment, with a slug and a public page. |
| Provider preferences | The provider object on a request: order, only, ignore, sort, max_price, require_parameters, data_collection, zdr, … |
| Quantization | The numeric format weights are served in (int2 … fp32). Distinguishes endpoints of the same model; filterable. |
| Rankings | Models by tokens processed over the last 30 days. Probe traffic counts, like any billed request. |
| Routing | Choosing which endpoint serves a request: capacity, then health, then a price-weighted balance — steerable end to end. |
| Routing suffix | :free, :nitro, :floor — appended to a model id to change how the endpoint is chosen. |
| Spend limit | An optional USD ceiling at key, workspace or account scope, with an optional UTC reset window. Breaching answers 402. |
| Statement | A provider's closed month: gross, platform share, net, generation count; the invoice a payout settles. It claims the earnings nothing had claimed when the month closed, so a late arrival is paid by the next one. |
| Source | A provider origin Heimdallm pulls the model list from and sends inference to. |
| Statement statuses | open, finalized, held, paid. A paid statement carries its payout reference. |
| Throughput | Completion tokens per second during generation, measured over the last day on streamed generations. |
| Time to first token | Seconds from request to first streamed token. A streamed-only figure; the number interactive use feels. |
| Token | The unit text is metered in, for context and for price. Embeddings and transcriptions bill in their own units. |
| Uptime | Share of requests that succeeded over a window (1 day, 7 days). |
| Wallet binding | A free one-time signature matching a wallet to an account so its deposits credit automatically. |
| Workspace | A compartment inside an account: its own keys, spend limit and routing defaults. Keys authenticate to exactly one. |
| ZDR | Zero data retention. Declared per endpoint and reviewed once per provider by the marketplace; zdr: true routes only to a declaring endpoint whose provider is approved. Undeclared, pending and denied never qualify. |