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 |
|---|---|---|---|---|
jers | chosen by its language router | reads the script and language of the state, then picks English or multilingual; typed-decisions only when named | ||
jers-latest | alias of jers | pin a checkpoint name before relying on exact behaviour | ||
jers-english | ModernBERT-large, 421M parameters | 512 tokens | about 475 / 330 tokens | the most measured here; the engine's authors report 0.78 on MASSIVE intent in English (0.86 on English XNLI) |
jers-multilingual | mmBERT-base, 322M, 100+ languages | 1024 tokens | about 990 / 780 tokens | the engine's authors: +21 points over English on non-English XNLI; Swedish 0.57 against English 0.78 |
jers-typed-decisions | ModernBERT-large fine-tuned on four workflows | 1024 tokens | about 990 / 790 tokens | 0.77 on its own workflows (customer service, invoices, security incidents, agent traces), where the base checkpoints score 0.34 to 0.36 against a 0.46 majority baseline; best for multi-question workflow decisions like those |
jers-ft-<tenant>-<date> | fine-tuned on one tenant's golden cases (jers_finetune.py) | as its base | as its base | 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); accuracy figures are the engine's authors' unless marked. The English checkpoint answers a three-question request in 22 to 25 ms of engine time on this Mac (medians of 20 requests, three runs, 2026-09-23).
default, auto and jers-latest are aliases of jers. An unknown name is a 422 that lists these. Every answer carries model (the model name with aliases resolved: jers-latest, default and auto come back as jers) and engine (the engine's name, the checkpoint that answered, and its metadata), so you always know what produced a number. The engine's authors report that the English checkpoint collapses outside English while staying confident, which is why jers routes by language before the engine reads anything: name jers-english only for English states.
Versions
GET /v1/models reports the runtime library version and the loaded checkpoints; treat jers-* names as "the checkpoint Jers runs today", record engine from the answer next to any measurement, and run your golden set (Quality) before and after the engine is upgraded.