Bespoke Nimble

Bespoke Nimble is a repository from Bespoke Labs that provides a model, training data, and a training recipe for making typed decisions about text. Given a text and a flat schema of enum or boolean questions, it returns the chosen answer and the probability of each allowed answer. It can run locally on Apple Silicon Macs or NVIDIA GPUs, and a hosted API is also referenced.

What It Is

Nimble is a decision model in the style of TypeSafe's Jev. It makes each decision in a single step without writing out reasoning first. The scorer reads the model's logits for one-token answer codes and converts them to probabilities with softmax, so there is no generated JSON to parse. The released Bespoke-Nimble-9B is a LoRA fine-tune of Qwen3.5-9B trained on the answer tokens only.

Typical tasks include routing a request to a destination, checking a yes/no condition, applying a policy to get an allowed outcome, and rating an outcome on an ordered scale.

How It Works

You supply a context and a schema where each field is an enum (1 to 255 choices in the latest release) or a boolean. On a Mac, ParallelScorer processes the shared context once and scores all fields in parallel; the CUDA scorer scores each field with the full prompt. Fields are scored independently, so one field cannot see another's answer. Prompts are limited to 8,192 tokens in the latest release.

Documented limits: text-only input, no free-text generation or nested JSON, and probabilities are not a guarantee of correctness, so thresholds should be tested on your own data.

Training Data and Methodology

The repo documents a contrastive data curation method: paired examples differ in one fact so the correct label flips, and both examples are verified with separate model checks before labels are applied by code. The published dataset has 2,676 training and 324 held-out examples across ten subject categories, with synthetic labels that no person has reviewed. The README reports 90.1% agreement with reference labels on the held-out set, versus 66.4% for the base model and 93.2% for Jev 1.13.0.

Setup Path

Use Python 3.12, download and merge the model weights, then load either the MLX or CUDA scorer and call score with a text and schema. Unquantized 9B weights take about 18 GB. The README also describes an agent skill for building on the hosted Nimble API, installable for Claude Code or via skills.sh, plus a Modal/SGLang serving guide and public benchmark suite.

Current hosted API and pricing

Official API pricing is $0.04 per million input tokens, with output tokens free. Shared text is billed once per request. The docs list nimble-v4-max and nimble-latest; nimble-v4 is deprecated and aliases v4-max at the same rate, and v3 is deprecated. This hosted service is distinct from the local 9B repository materials. A verified new account receives a one-time $1 credit. The console currently says new signups are paused and offers a waitlist.

The latest local checkpoint uses temperature 1.0 without a separate fit; older calibrated revisions have different settings. Public repository access is not a blanket software or model license.