Built in the open. One contract at a time.

Typed decisions, with a contract

A dedicated evaluation surface takes a state and 1–64 named questions. It preserves stable answer keys, declared types, and usage metadata.

Vectors for retrieval

The reference includes twelve embedding models, native dimensions, language hints, float or base64 output, and batches up to 64 items.

A second pass for relevance

Five rerank records cover up to 100 candidate documents, returning original input indexes and model-specific scores.

PaymentsDevelopment reference

Beyond a text box

Images, image edits, video jobs, speech, transcription, files, and expiring artifacts each need their own contract.

Review the payment before signing

A wallet can identify the requester and authorize a maximum for a metered inference operation.

Make every billing unit explicit

The archived pricing design separates input, output, cache reads, minimum charges, and modality-specific units.

Quote, authorize, reconcile

A bounded request requires an immutable rate snapshot, a maximum authorization, and separate actual-cost settlement.

AccountsDevelopment reference

Keep cache usage visible

Some models report cached input separately from ordinary input; a missing cache field is unknown rather than zero.

Grants with an audit trail

The source account surface includes redeemable grants, expiry, remaining balance, and a ledger entry for each redemption.

AccountsDevelopment reference

Entitlements belong to the account

The account reference distinguishes shared account limits from tighter per-key scopes, budgets, concurrency, and expiry.

Reconcile model-family usage

Provider-specific usage formats can expose input, output, cache reads, and cache creation through different fields.

Qualifying a new model generation

The archived catalog contains a Claude Opus 5 reference with a 1M-token context, 128K output, and explicit reasoning levels.

Version rates alongside profiles

The reference contains pinned and rolling DeepSeek variants with different capabilities and cost parameters.

Long context needs bounded admission

The exported GLM 5.3 record advertises 1M context, 128K output, and a defined reasoning list.

Qualify each input modality

The archived Flash record includes text, image, and video inputs with a 1M context window.

Reasoning is a model capability

The imported catalog includes a GPT 6 Astra specification with reasoning choices and multimodal input metadata.

Keep multimodal quotes reproducible

The Kimi K3 record combines a 1M context, text/image/video input, and a model-specific reasoning list.

Model families are not interchangeable

The archive includes distinct Opus and Sonnet records, limits, rates, and supported reasoning levels.

Designing a short promotion

The source journal described a short, broadly applied discount. It is retained here as a campaign design topic, not an active Noemad offer.

A campaign should expire predictably

A second archived topic covered a 48-hour discount window. The historical date in this route is a reference identifier, not a Noemad release date.

Model-specific grants and caps

The source campaign used a model-specific grant with a bounded promotional allocation. No grant or free allowance is active in this preview.

Discounts for an eligible subset

This archived topic demonstrates a campaign restricted to a selected model tier.

Separate grants for separate models

The source offered separate promotional pools for two model families. This is an entitlement design scenario only.

A bounded introductory allowance

The imported topic covered an introductory allowance for a particular cost tier and associated coding model.

Shared campaign capacity

A shared promotional pool requires coordination across participants and requests.

Scheduled access windows

The archived topic used a recurring time window for an eligible model.