Bespoke Nimble 9B
Bespoke Nimble 9B is an Apache-2.0 PEFT/LoRA adapter for Qwen3.5-9B, not a standalone 9B checkpoint. It scores enum, boolean, and rubric answers directly, supports up to 255 choices per field and an 8K request limit, and requires the separate Qwen3.5-9B base weights. Its narrow author-reported evaluation should be recalibrated on the target workflow.
Deployment and license note
Decision models return bounded choices, scores, or yes/no probabilities instead of long-form text. Public benchmark results are useful for screening candidates, but production thresholds should be calibrated on a labelled set from your own routing, guardrail, or action-selection workflow.
Typed-decision adapter
What this model actually is
Bespoke Nimble 9B is a PEFT/LoRA adapter for Qwen3.5-9B, not a standalone 9B checkpoint. The repository contains the adapter, tokenizer, prompt builder, and reference inference code; the Qwen3.5-9B base weights must be downloaded separately.
Decision shortcut: Use it for bounded choices, booleans, rubric levels, routing, and guardrail decisions where direct probabilities are more useful than free-form prose. Do not choose it as a general chat or coding model, and calibrate thresholds on a labelled set from the real workflow.
The small adapter download does not remove the memory required by the full 9B base model.
It scores allowed answer tokens directly instead of generating a normal assistant response.
The current repository documents these limits for the updated checkpoint.
This is a narrow author-reported result and should not be compared directly with general LLM benchmarks.
Good reasons to choose it
- Routing a request into a fixed set of workflows
- Policy or guardrail checks with an explicit human-review threshold
- Scoring a small rubric when probability and calibration matter
Verify before deployment
- Download the exact Qwen3.5-9B base revision expected by the adapter.
- Use the repository prompt builder and PEFT loading path instead of treating the adapter as a GGUF chat model.
- Measure false-positive, false-negative, abstention, and calibration behavior on a private golden set before automation.
Task Fit
Bounded routing, classification, scoring, guardrails, and action selection with structured probabilities.
Not marked for code agent in the current library.
Not marked for code in the current library.
Not marked for chat in the current library.
Not marked for rag in the current library.
Not marked for vision in the current library.
Not marked for image generation in the current library.
Not marked for video generation in the current library.
Not marked for voice in the current library.
Evidence level
How this is assessed →Memory and hardware-fit values are planning estimates unless a tested report is linked.
Variants and Quant Artifacts
Choose the artifact first; hardware fit follows from RAM, VRAM, format, and runtime.
| Quant | Format | Quality | Min RAM | Reco RAM | Runtime | Action |
|---|---|---|---|---|---|---|
| BF16 + LoRA | peft-adapter | high | 24GB | 32GB | transformers | Plan with this |
Recommended Hardware
Lowest estimated 5-year cost that can run this model.
Enough unified/system memory with a balanced 5-year cost.
Highest local performance signal among compatible hardware.
Benchmarks
Sources and compatibility
Execution evidence
Run Bespoke Nimble 9B with a documented recipe
Recipes connect hardware, a model artifact, tools, settings, verification, and a reportable result.
No verified recipe is linked to this record yet.
Compatibility estimates remain available in the planner. A recipe appears here only after its exact stack and verification protocol are documented.