Bespoke-Nimble-9B
An open Jev-style LoRA on Qwen3.5-9B from Bespoke Labs, trained on 2,676 contrastively curated examples to score the allowed answer tokens directly for enums, booleans and rubric levels. Recipe, data and a public benchmark suite are released with it.
On Bespoke's own 324-example held-out set Nimble scores 90.1% (Jev 93.2%) with an expected calibration error of 0.054 after the 22 September temperature fit; across 13 human-labelled public subsets it reaches 74.8% macro (Jev 76.0%). The 24 September checkpoint serves 8,192 tokens and 255 choices and ships with temperature 1.0; the earlier checkpoint is revision original-2676. The repository holds the adapter (about 165 MiB), not the base weights. A public demo without a key runs on Modal and speaks /v1/systemone; median latency there is 106 ms on an H100.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one calibrated probability
- score — places the input on an ordered scale
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 9B |
| Base model | qwen/qwen3.5-9b |
| Maker | Bespoke Labs |
| Released | 2026-09-18 |
| License | apache-2.0 |
| Reported accuracy | 90.1% |
| Reported latency | 106 ms median on H100 |
Hosted API
Served by Bespoke Labs. Get access · API docs.
Get the weights
pip install systemonemodels
systemone pull bespoke-labs/bespoke-nimble-9b
The files are served from the maker's Hugging Face repository, bespokelabs/Bespoke-Nimble-9B, and verified against the checksums recorded here.