Bespoke Nimble
A 9B local typed-decision adapter for flat enum, boolean, and rubric decisions over text.
Bespoke Labs publishes model artifacts, reference scorers, training data notes, and evaluation guidance for Bespoke-Nimble-9B.
- Provider
- Bespoke Labs
- Reviewed
- September 23, 2026
- Project status
- Official
- Verification
- Verified
- Categories
- automation, research
- Compatibility
- Python, Hugging Face, CUDA, Apple Silicon
Inputs
- ✓Text context
- ✓Flat enum or boolean schema
Outputs
- ✓Selected answers
- ✓Candidate probabilities
- ✓Typed fields
Interface
- ✓structured-output
Decision tasks
- ✓Classify
- ✓Choose
- ✓Score
- ✓Route
Verified model facts
Provider facts and directory notes are separated. Time-sensitive claims include attribution and should be checked again before production use.
- Current verified version
- Bespoke-Nimble-9B revision 93ec5d6
- Model family
- Bespoke Nimble
- Release date
- Unknown
- Access methods
- hosted-api, local-package, open-weights
- Hosted API
- Yes
- Local deployment
- Yes
- Downloadable weights
- Yes
- License
- Apache-2.0
- Supported languages
- English
Commercial-use note
The published adapter is Apache-2.0. The required Qwen3.5-9B base checkpoint and other dependencies retain their own terms.
Context or input limits
Training used prompts up to 2,048 tokens. The hosted path accepts up to 8,192 tokens per question, while the project says shorter prompts are better tested.
Pricing
Local software pricing is not applicable; hardware cost varies. Hosted API pricing was not confirmed in the reviewed sources.
Latency and rate limits
The project publishes hardware-specific measurements for its own evaluation runs. They are not directly comparable across providers and were not reproduced here.
Strengths
- ✓Publishes model artifacts, reference scoring code, and training notes.
- ✓Supports local execution on documented Apple Silicon and NVIDIA paths.
- ✓Returns candidate probabilities without free-form generation.
Limitations
- ✓Text-only input and flat enum or boolean schemas.
- ✓The project describes rough edges and narrow curated training domains.
- ✓Local deployment requires substantial memory and compatible hardware.
Safety notes
- ✓Probabilities are not correctness guarantees.
- ✓Retest every threshold on local data and include a no-match or review path.
Implementation guidance
Keep the declared inputs and outputs visible in code review. Validate every model response before it reaches a database, API, or customer-facing action.