Bespoke Labs (@madiator / @bespokelabsai) open-sourced Bespoke Nimble: an open-data, open-recipe, open-weight Open Jev-style decision model. The key idea is contrastive data curation—slightly changing focal facts to flip labels—so the model learns to discriminate, with implicit calibration and no probability labels. Training is a LoRA finetune of Qwen3.5-9B without distilling Jev (Jev is eval-only) and without RL yet; serving uses parallel constrained decoding over candidate-token logits. On their curated eval, accuracy rose from ~66% (base Qwen) to ~90% (Nimble) vs ~93% for Jev, with ~100ms on an H100 and local MacBook runs. Code and weights are on GitHub and Hugging Face.

Key Takeaways

  • ✓Open-data/recipe/weights Open Jev via contrastive curation that flips labels on tiny fact edits; calibration is implicit without probability labels.
  • ✓LoRA on Qwen3.5-9B, no Jev distillation, no RL yet; parallel constrained decoding over candidate-token logits.
  • ✓Curated eval ~66%→90% (Jev≈93%); ~100ms on H100 and runnable on a MacBook; GitHub+HF released.
ADSponsored