Bespoke Labs released Nimble, a model for typed decisions on text. It answers schema-defined questions in one step without reasoning. The model is trained on Qwen3.5-9B with LoRA on answer tokens only. It achieves 90.1% accuracy on 324 held-out examples. The latest checkpoint supports 8,192-token context and up to 255 choices per field. It runs on Mac with Apple Silicon or Linux with an NVIDIA GPU. Training data and examples were published on September 20, 2026. A public benchmark suite compares Nimble and Jev on 13 datasets. The hosted API prompt limit was raised to 8,192 tokens on September 19, 2026. A temperature fit was added on September 22, 2026 to align probabilities with correctness.
What changed
Nimble previously used a 2,048-token prompt limit and lacked public training data.
What it unlocks
Making typed decisions from text without intermediate reasoning steps.
- 90.1% accuracy on 324 held-out examples
- 8,192-token context limit
- 255 choices per field
- 2,676 training examples
What you need to act on it
- Apple Silicon or NVIDIA GPU with BF16 support
- Python 3.12
- 18 GB+ memory for unquantized weights
Sources