What Solar Decide is
Solar Decide is Upstage’s System One model. You send a state, which is the text to judge, and one or more typed questions. It returns one answer per question and writes no prose. Upstage serves it on Solar Mini 4, a 35B-parameter mixture-of-experts model with 3B parameters active per token. It uses the same /v1/systemone request schema as Jev, so a client written for one can point at the other.
What it returns
The three question types match Jev’s. A Choice returns the best option, the probability of every option and a confidence value. A Score returns an expected value from 0 to the number of levels minus one, with the probability of each level. A Noul returns one number, the probability that the answer is yes.
Each question takes one forward pass and output tokens are not billed. A Choice takes 2 to 26 options. To decide among more, Upstage says to split the decision into several questions or narrow the candidates first.
What it is good at
The context window is 512K tokens, so a whole contract or report can be the state. Solar Mini 4 is built for Korean, English and Japanese, and Upstage’s own description stresses its Korean. OpenRouter’s listing names routing, classification and policy checks as the targets. Those map to intent and model routing, support inbox triage and compliance verification.
Upstage has published no accuracy figures in the docs checked on 30 September 2026. The only measurements found are third-party. In pull request 17 to zero-shot-ie-bench, its author reports 95.8% on a 48-question sentiment and topic pool against 93.8% for Jev, and 100% on 54 texts across nine languages, including Sinhala, Icelandic and Welsh. Those are small samples from one person’s test set.
What it is not for
It cannot write text, explain an answer or handle more than 26 options in one Choice. It is a beta, and the schema may change. The same pull request found that a Noul question must leave out the criteria key, because an explicit null is rejected.
Access
Solar Decide has been in beta on the Upstage Console since 22 September 2026. Call POST https://api.upstage.ai/v1/systemone with an Upstage API key. OpenRouter added it on 28 September as a decisions model, not a chat model. The model page lists the Upstage Console as the only platform, and OpenRouter shows no Hugging Face repository for it, so there are no downloadable weights.
Specifications
| Question types | ChoiceScoreNoul |
| Max Choice options | 26 |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 524,288 tokens |
| State budget | Not documented |
| Rate limit | Set by commitment tier: 100 requests and 250,000 tokens per minute at Tier 0, up to 8,000 requests and 3,000,000 tokens per minute at Tier 4. Legacy tiers are capped at 100 requests and 50,000 tokens per minute. |
| Endpoint | POST https://api.upstage.ai/v1/systemone |
| SDKs |
Upstage lists a 512K context, which OpenRouter gives as 524,288 tokens. A Choice question takes 2 to 26 options, because the endpoint reads the answer from single-token labels A through Z. A longer list returns a 422 error. Upstage does not document a maximum number of Score levels, a maximum number of questions per call, or a separate limit on state tokens.
Versions
- solar-decide, 22 Sep 2026, The only version Upstage lists, marked Beta. Upstage warns that the /v1/systemone schema may change before general availability. OpenRouter serves it as upstage/solar-decide, snapshot solar-decide-20260928. Release notes
Use cases
What people use Solar Decide for, one page per pattern.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Compliance verification with System One models
A compliance review is a fixed checklist run against a changing document. Jev answers the whole checklist in one request, returning a probability, a label or a rated level per item, so code can clear the clear-cut findings and send the borderline ones to a person.