Bespoke-Nimble-9B

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

• Maturity4/10

  Surprisingly complete open recipe (data/train/serve/evals) for a 1–2 day build; still one-epoch LoRA, no hosted SLA, no RLCD.

• Capability6/10

  Correct System One shape (choice / boolean / rubric score via candidate logits); close to Jev on author holdout and public human suite, weaker calibration (ECE).

• Adoption12/100

  Strong X launch buzz (~176k impressions) but D+1 awareness; ~161 GitHub stars and low HF downloads — early niche, not category default.

Vendor claims

• ~100ms median on H100 for a scoring subset (author timing; not full holdout)[Vendor claim — not independently verified]
• Holdout accuracy 90.1% vs base Qwen 66.4% vs Jev 1.13 93.2% (author 324-example set)[Vendor claim — not independently verified]
• Public human-label suite ~74.8% vs Jev ~76.0% (author-reported)[Vendor claim — not independently verified]
• Did not distill from Jev — contrastive curation + LoRA on answer tokens only[Vendor claim — not independently verified]

Bespoke-Nimble-9B is not a frontier base model — it is a LoRA adapter (PEFT r=16, ~165 MiB) on Qwen3.5-9B, shipped by Bespoke Labs as an open data / open model / open recipe for an open Jev.

Given text + a flat schema, it returns typed answers and probabilities over allowed tokens. It does not generate free-form reasoning as the primary output. Inspired by TypeSafe’s System One / Jev; the authors state they did not distill from Jev.

What it is

• Adapter: bespokelabs/Bespoke-Nimble-9B (Apache-2.0; needs Qwen/Qwen3.5-9B base)
• Recipe + local scoring: github.com/bespokelabsai/nimble
• Schema fields: boolean, enum (≤26 choices), rubric scores as ordered enum strings; returns selected value + per-option probabilities
• Method highlight: contrastive data curation (flip one fact → negative) + LoRA on answer tokens only
• Launch note: Mahesh Sathiamoorthy (@madiator) (18 Sep 2026)

See also catalog peers: Kev (smaller LoRA reconstruction), Jev (hosted reference).

Limits (read before you ship)

• Built in ~1–2 days — expect rough edges; research prototype, not a hosted product
• Max 26 choices per field; prompts over 2,048 tokens are rejected
• Author holdout is synthetic / model-checked (324 examples); public suite labels are human but criteria wording is a free parameter
• Probabilities are softmax over candidates at temperature 1.0 — not claimed as temperature-tuned calibration; ECE trails Jev on the public suite (author docs)
• Merged Mac path needs substantial RAM (~18 GB+ BF16 base); quality checks emphasized on the adapter CUDA path
• Do not use for production decisions that affect people without your own labelled eval

Why it is in the catalog

Public typed decision I/O, open weights + full recipe, and explicit System One framing. Plotted next to Kev as an open LoRA point — with artifactKind: lora-adapter so the card does not pretend this is a standalone frontier model.