Nimble returns typed decisions, not chat text

On 2026-09-24, Bespoke Labs described an updated checkpoint of its Nimble model in the project repository on github.com. Nimble takes text and a schema and makes typed decisions about that text. The schema is the list of questions to answer. Each question is either a choice from a list you supply or a true or false question. For each question, the model returns the answer it picked and the probability of each allowed answer.

What Nimble does and does not do

Nimble makes each decision in one step and does not write out reasoning first, which the repository describes as fast. It only picks from the answers you supply. It cannot write text of its own, such as an explanation, and it cannot return nested JSON or a passage taken from the context. The schema must be flat: each field is an enum with a fixed list of string choices, or a boolean. In the latest release an enum field can hold up to 255 choices, and each prompt can be at most 8,192 tokens, including the schema and the field name. Fields are scored separately, so one field cannot see another field's answer, and your code has to check that the answers are consistent.

The repository states plainly that the probabilities are not a guarantee that an answer is correct, and that they are scaled so that they add up across the answers you supplied. If none of your answers may fit, add an answer that means "no match", and test any probability threshold on your own data.

Where this touches office work

The fit is with work that already has a fixed set of outcomes. Document Desk reads incoming invoices, tenders and contracts and enters them into business systems; a schema of enum fields could carry a supplier match, a document type or a routing destination, with the probability shown next to each pick. Approval Runner checks requests and routes them for sign-off, which is close to the routing and condition-checking tasks the repository describes. Company Brain answers from a company's own documents, and Catalogue Builder turns inconsistent supplier files into product records; both depend on decisions that can be expressed as choices rather than free text.

A team would want to try this on a small set of its own invoices or requests first: define the fields, include a "no match" answer, and compare the picked answers against what staff decided. The model runs on a Mac with Apple Silicon or on a machine with an NVIDIA GPU, so a pilot can stay on local hardware.