Ollaya 0.17.0 ships coba, the first decision model we trained ourselves. Like every decision model, it reads a state and typed questions (yes/no, choice, score) and answers each with probabilities in a single pass, without generating text.
coba is an encoder, so it is fast: coba:400m answers five questions in 24 ms on an RTX 5090. LLM-based decision models around 0.8B take 60–69 ms. Its ternary sizes are small enough for a browser, and ollaya.dev/try runs coba:150m-ternary in the page: what you type never leaves it.
Where it stands
On Ollaya Bench 2026.10, coba:400m has the best typed-decisions score of any encoder under 1B parameters not trained on that dataset, and coba:1b is higher still. It is not the most accurate decision model overall: LLM-based models around 0.8B score higher on the whole benchmark, though not on typed decisions.
| Model | Overall score | Typed decisions | Five questions, RTX 5090 / CPU |
|---|---|---|---|
coba:1b | 36.3 | 42.6 | 21 / 1530 ms |
coba:400m | 35.8 | 40.1 | 24 / 919 ms |
coba:150m-ternary, in-domain | – | 31.9 | 17 / 496 ms |
coba:68m-ternary, in-domain | – | 24.2 | 15 / 323 ms |
laya:en | 30.1 | 5.9 | 18 / 945 ms |
decider:0.8b, an LLM | 47.8 | 27.6 | 60 / 3106 ms |
Scores are chance-corrected accuracy: 0 is chance, 100 is every answer right. The overall score is the mean over 14 tasks, 13 public ones and typed decisions. The ternary tags are in-domain: they were trained on the train splits of most public tasks, so they have no comparable overall score. Typed decisions is not among those tasks. CPU times are on a Threadripper 3960X. Every number, with its raw data, is on Ollaya Bench and the coba page.
How it is made
coba runs Laya's decision network, which Ollaya already ran, on Ettin encoders. It learned from pplx-decider:flash's probabilities, on permissively licensed data only, and never saw typed decisions' test set. The ternary tags are trained quantization-aware from the full-precision model of the same size. coba is English only.
Try it
- In your browser: ollaya.dev/try
- On your machine, with Ollaya 0.17.0:
ollaya run coba:400m --ask "Is the customer asking for a refund?" "I was charged twice for order A-104." - In your own page:
npm install @ollaya/webruns the ternary tags with onnxruntime-web. - The weights: huggingface.co/ollaya-dev
Next
sura is Ollaya's decoder family, built on Qwen3.5 and distilled from pplx-decider, for when accuracy matters more than speed. It is in training: sura-0.8b and sura-2b.