Bespoke Nimble
Nimble (github.
Overview
Nimble (github.com/bespokelabsai/nimble) from Bespoke Labs ships contrastive data curation, LoRA training on answer tokens only, and serving for Bespoke-Nimble-9B on Hugging Face. You pass text plus a schema of Choice lists or true/false fields; the model scores one answer token per question in parallel without chain-of-thought prose. README reports 90.1% agreement with reference labels on 324 held-out examples versus 66.4% for the Qwen3.5-9B base and 93.2% for Jev 1.13.0, with explicit note that Bespoke did not distill from TypeSafe. September 2026 temperature fitting improves probability calibration while keeping discrete picks stable. Runs on Mac Metal or Linux CUDA with documented merge steps for LoRA adapters.
Problem: Teams want locally runnable Jev-style classifiers with published data curation and training recipes, not opaque distillation from a hosted gate.
Built for: Applied researchers on Apple Silicon or NVIDIA who need Choice and boolean fields from a schema with probabilities, plus scripts to reproduce Bespoke-Nimble-9B.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Bespoke Labs
- Organization
- Bespoke Labs
- Handle
- @bespokelabsai
How Jev is used
- Role in the product flow
- Local schema-driven Choice and Noul classification from merged 9B weights
- Primitives
- ChoiceNoul
- State in
- Prompt text up to 2,048 tokens including schema field names plus per-field candidate lists or boolean questions defined in nimble/scoring/parallel_schema.py.
- Decision out
- Selected option per field with normalized probabilities across supplied answers; no free-text completions.
- Validate schema against model contract and prompt hash
- Encode prompt once per request
- Score allowed answer tokens in parallel for each schema field
- Return picks and probabilities for downstream thresholds
Nimble is the teach-the-recipe counterpart to hosted Simple Jev or SemIf servers. Operators define explicit enums and booleans in code, call local inference, and treat probabilities as hints that still need domain calibration. The repo foregrounds data edits that flip labels under contrastive curation, which is rare in integration listings that only wrap API keys. Because fields cannot depend on each other, your orchestration layer must enforce cross-field consistency after Nimble returns independent answers. Pair with classifier.dev when you want managed latency without merging nine billion parameters on a laptop.
Sourced performance claims
- README cites 90.1% label match on 324 held-out examples for Bespoke-Nimble-9B versus 93.2% for Jev 1.13.0.Source: github.com/bespokelabsai/nimble README
- Training set documents 2,676 curated examples with open curation and training scripts in-repo.Source: github.com/bespokelabsai/nimble README
- Public GitHub repo bespokelabsai/nimble had about one thousand six hundred seventy eight stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Open data curation, training, and MLX or CUDA serving paths
- Hugging Face Bespoke-Nimble-9B with schema contract files
- Parallel answer-token scoring inspired by System One Choice docs
- Documented limits: text-only, max 26 enum strings per field
Stack
- Python
- PyTorch
- MLX on Apple Silicon
- LoRA on Qwen3.5-9B
- Hugging Face Hub
Pricing: Open source recipe; you fund GPUs, RAM for merge steps, and any cloud training you run.
Links
FAQ
- Did Bespoke distill from TypeSafe Jev?
- README states they did not distill from Jev; the repo shares curation, training, and serving so others can research similar models.
- Which primitives does Nimble support?
- Schema fields are Choice over explicit lists or boolean (Noul-style) questions. There is no free-form text or nested JSON output.
- Can I trust default probabilities in production?
- README warns probabilities are calibrated on their curated set and still need threshold tuning on your data, especially when no supplied answer fits.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.