Bespoke Nimble
Open data, model, and recipe for typed decisions without Jev distillation — LoRA on Qwen3.5-9B with contrastive curation, scoring answer tokens via logits.
Bespoke Labs released Bespoke Nimble: an open-source path to Jev-style typed decisions — data curation recipe, model weights, training code, and serving stack, all public.
Built in about two days on Qwen3.5-9B with a light LoRA fine-tune on 2,676 curated examples. No distillation from Jev; Jev is used only as an evaluation reference. On their 324-example held-out set, Bespoke-Nimble-9B hits 90.12% vs 93.21% for Jev 1.13.0 and 66.4% for the base Qwen model.
Contrastive data curation
The core training trick: build pairs of nearly identical prompts where one key fact flips the correct answer. The model learns which evidence actually drives the decision — not spurious correlations.
How it serves
Nimble does not generate CoT or JSON. It reads logits for one-token answer codes, softmaxes them into probabilities, and your code assembles structured output. On Mac, ParallelScorer shares context once and scores fields in parallel.
What you can build
- Route requests to destinations
- Check yes/no conditions on evidence
- Apply policy rules with typed outcomes
- Rate outcomes on ordered scales
Run it
git clone https://github.com/bespokelabsai/nimble.git
cd nimble
# See README for Mac (Apple Silicon) or CUDA setup
pip install -e .
Model weights: Bespoke-Nimble-9B on Hugging Face.