Biplov GautamOn 15 September 2026 TypeSafe AI shipped Jev and gave the category a name: a System One model takes a state and a set of typed questions and returns one answer per question with a calibrated probability, generating no text. Ten days later there are fifteen families you can use, and they disagree about almost everything except that shape. This is the map. Every model here has a page on the registry with its manifest, its files and its reported numbers.
Three primitives
Every family answers the same three question types, in TypeSafe's vocabulary:
- Choice: pick one option from a set (up to 255 in most implementations).
- Score: place the input on an ordered scale of 2–10 levels.
- Noul: a yes/no question answered with a single probability.
Some add more. GLiNER2.5-Decide extracts spans and relations; CLM exposes a
ranking endpoint; the registry's manifest records all of these as
capabilities, so you can filter by what a model decides.
Hosted, open, or both
| Model | Maker | Params | Weights | API | Licence |
|---|---|---|---|---|---|
| Jev 1.13 | TypeSafe AI | undisclosed | no | $0.042/MTok in, output free | proprietary |
| Tev1-4B | Together AI | 4B | on HF, licence pending | Together, early access | unspecified |
| GLiNER2.5-Decide | Fastino Labs | 340M | yes | Fastino | Apache-2.0 |
| Nimble-9B | Bespoke Labs | 9B (LoRA) | yes | public demo | Apache-2.0 |
| Kev | Jared Palmer | 0.8B–27B (LoRA) | yes | self-host | Apache-2.0 |
| Laya | Convai Innovations | 421M | yes | self-host | Apache-2.0 |
| CLM v0.1-8B | Contrastive-LM | 8B (heads only) | yes | self-host | Apache-2.0 |
| Decider | Mapika | 0.8B–35B | yes | self-host | Apache-2.0 |
| Von 1.2 | wfzyx | 395M | yes | self-host | Apache-2.0 |
| JevK5 | Alibi Serikbay | 2B–9B | yes | self-host | Apache-2.0 |
| OpenThai-SystemOne | iApp | 0.8B | yes | free hosted | Apache-2.0 |
| cua-s1-forms | Cua | 706K | yes | self-host | MIT |
| NanoJev | TianyuCodings | 0.6B | yes | self-host | MIT |
| Jev-Omni | akhilaaa3 | 12B | yes | self-host | Apache-2.0 (Gemma base) |
| Qwen3.5-4B scorer | pngwn | 4B (LoRA) | yes | self-host | CC-BY-NC-4.0 |
Two are commercial. Jev is closed and served only through TypeSafe's API. Tev1's weights are on Hugging Face, but without a licence until Together finalises one, so the registry links to them rather than mirroring them.
Two ways to build one
The families split by where the decision comes from.
Encoders with a decision head. Laya, Von and GLiNER2.5-Decide put a bidirectional encoder (ModernBERT-large, DeBERTa-v3-large) over the state and the options in one sequence and score a marker per option. They are small, run on a CPU, and answer in tens of milliseconds: Laya reports 33–39 ms on a T4, GLiNER2.5-Decide 38 ms on a V100.
Decoder backbones with a readout. Kev, Nimble, Decider, JevK5, Tev1, NanoJev, OpenThai and the pngwn scorer start from a Qwen (or Gemma) language model and replace generation with a readout: a pointer head, a softmax over the answer-letter logits, a 256-way slot head. They are bigger and want a GPU, but inherit what the backbone knows about the world. Decider reports a 3.2 ms median with CUDA graphs; JevK5 13 ms on an H100.
CLM is the odd one out: two contrastive heads on frozen Qwen3-8B embeddings that score a state against candidate actions with InfoNCE, so a decision is a nearest neighbour search rather than a classification.
What was actually measured
No two makers use the same benchmark, so the table on each model page names the suite next to the number. The ones that overlap:
- JevBench public hard tier: JevK5 v0.3 78.4%, Decider-4b 64.9%, Decider-2b 57.7%, Von 36.9%.
- Bespoke's 13 public subsets (macro): Jev 76.0%, Nimble-9B 74.8%, OpenThai 74.3%.
- Fastino's fast-decisions suite: GLiNER2.5-Decide 60.2%, JevK5 57.6%.
Calibration is where the field is honest in an unusual way. Kev, Nimble,
Decider, Von, JevK5 and Jev-Omni all publish an expected calibration error
(0.035–0.054), and several say plainly when a temperature was fitted on the
test set. The registry shows calibration_error beside accuracy on every page
for that reason: a decision you cannot threshold is not much of a decision.
One API shape
Nearly every open family serves TypeSafe's POST /v1/systemone contract, so a
client written for Jev works against Kev, Laya, CLM, Decider, Von, JevK5,
OpenThai or Nimble by changing the base URL. That is the field's quiet
agreement, and it is why the registry records runtime and hosted per model
rather than inventing another format.
Where this list comes from
Facts on this page and on the model pages were taken from each maker's model card, README, launch post or API documentation on 25 September 2026, with the Hub's own file sizes and checksums. Where sources disagree, the model page says so. If you make one of these models, claim your organisation and the page is yours to correct.