Models
GET /v1/models lists what this gateway offers, whether each engine answers right now, and what is pinned.
| Name | Checkpoint | Window | State read (short question / 10-20 options) | Notes |
|---|---|---|---|---|
jepela | chosen by its language router | reads the script and language of the state, then picks English or multilingual; typed-decisions only when named | ||
jepela-latest | alias of jepela | the same as jepela; no name keeps answers from changing (see Versions) | ||
jepela-english | English | 512 tokens | about 475 / 330 tokens | the most measured here |
jepela-multilingual | 100+ languages | 1024 tokens | about 990 / 780 tokens | for states that are not in English; less measured here |
jepela-typed-decisions | English, trained on four workflows | 1024 tokens | about 990 / 790 tokens | best for multi-question workflow decisions: customer service, invoices, security incidents, agent traces |
jepela-ft-<tenant>-<8 hex digits> | fine-tuned on one tenant's golden cases (see Fine-tuning) | as its base | as its base | the model that POST /v1/finetune gives: your tenant (as GET /v1/usage shows it) in lowercase letters, digits and hyphens, at most 30 characters, then the first 8 hex digits of the job id; listed by GET /v1/models only for that tenant, with "fine_tuned": true and its held-out accuracy; registered only when it beat its base on held-out cases |
Window figures measured on 2026-09-23 with this gateway's engine (see State). The English checkpoint answers a three-question request in 27 ms of engine time, timed on the engine's server (median of 25, 2026-09-24); the API's own work and the network between the caller, the API and the engine come on top.
Only the checkpoints served now are available in GET /v1/models. Naming one that is not served is a 422 that says which are. When jepela routes a state to a checkpoint that is not served now (a state that does not read as English while only the English checkpoint is served), the checkpoint that is served answers, never silently: the answer carries the warning language_not_served (see Warnings) and engine.checkpoint names the checkpoint that answered.
default, auto and jepela-latest are aliases of jepela. An unknown name is a 422 that lists these. Every answer carries model (the model name with aliases resolved: jepela-latest, default and auto come back as jepela) and engine (the engine's name, the checkpoint that answered, and its metadata), so you always know what produced a number. The English checkpoint collapses outside English while staying confident, which is why jepela routes by language before the engine reads anything: name jepela-english only for English states.
Versions
A model name chooses which checkpoint answers; it does not keep the answers from changing. jepela-english is always the English checkpoint, jepela always routes by language, and a jepela-ft- name is always the one model its job trained, until you delete it. Under the same name, Jepela's code around the checkpoint (how a state is read, matching, rounding, calibration) and the engine itself can change: Jepela's changes of 2026-09-29, for example, changed answers under unchanged names (probabilities rounded so that they sum to exactly 1, and a yes/no shift added to calibration fitted on your labels). No request field chooses or holds a version today.
GET /v1/models reports the engine's build and the loaded checkpoints: pinned.runtime is "engine" and the Jepela build id, the hash of the engine's code, the same id every answer carries as engine.runtime.version and GET /health gives as the engine's code. A new build id means the engine changed. Treat jepela-* names as "the checkpoint Jepela runs today", record engine from the answer next to any measurement, and run your golden set (Quality) again before and after the engine is upgraded, from time to time, and whenever your answers seem to have moved, so a change shows in your own numbers.