Bespoke-Nimble-9B

Bespoke-Nimble-9B

Scores yes or no questions, multiple-choice picks and rubric levels against a piece of text, using a LoRA adapter on Qwen3.5-9B. For developers who need a classifier that returns probabilities instead of free text.

Made by bespokelabs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Bespoke-Nimble-9B is a LoRA adapter for the Qwen3.5-9B model. You give it a block of text and a list of allowed answers, and it scores each answer directly. It does not write explanations or free-form replies.

It handles three kinds of fields: a choice from a list, a true or false value, and a numbered rubric level. The result includes the chosen answer and the probability of each option.

Who it suits: Developers who want a classifier that grounds its answer in text they supply, for example a policy check or a quality grade, and who have a CUDA GPU.

Industry | Who uses it | How they use it | Example
Retail | Support engineers | Check a customer case against a written policy | Is this item inside the return window
Content | Editors and reviewers | Grade a text against a rubric | Giving a draft a score from a small numbered scale
Software | Backend developers | Route a request to one of a set of labels | Choosing a category from a fixed list
Operations | Analysts | Turn a text record into a yes or no flag | Marking whether a record meets a stated rule

Anyone who wants chat answers, explanations or long generated text can skip it, since this adapter only scores the options you list.

Where to use it:

• Python on a machine with a CUDA GPU that supports BF16.
• Windows, macOS or Linux with Python, for downloading the files.
• Mac and other loaders are covered in the project's own quickstart.

Use case | Who | What they get back
Policy check | Support teams | A true or false answer with its probability
Rubric grading | Reviewers | A chosen level and an expected score weighted by probability
Label picking | Developers | One choice from up to 255 options with probabilities

How to set it up: You need the base model and a suitable GPU environment.

• A CUDA GPU with BF16 support
• A CUDA build of PyTorch, version 2.8.0 on CUDA 12.8 in the authors' own run
• Access to the Qwen3.5-9B base weights, which this repository does not include
• A Hugging Face token only if the repository is private

1. Download the repository into a local folder.
2. Install the pinned requirements from that folder.
3. Review the included helper before you import it. It does not need trust_remote_code.

pip install -U huggingface_hub
hf download bespokelabs/Bespoke-Nimble-9B --local-dir nimble-model
pip install -r nimble-model/requirements.txt

How to use it: Import the helper, load the model from the folder, then call score with a context and a schema describing the fields you want answered.

import sys
sys.path.insert(0, "nimble-model")
from inference import NimbleModel

model = NimbleModel("nimble-model")
result = model.score(
    context="The store accepts returns within 30 days. This item was bought 12 days ago.",
    schema={
        "eligible": {
            "type": "boolean",
            "description": "Is this item within the store return window?"
        }
    },
)
print(result["output"])
print(result["fields"]["eligible"]["probabilities"])

Tool | What it does
NimbleModel | Loads the adapter and its tokenizer from the local folder
score | Scores each field in your schema against the context
type enum with choices | Defines a field whose answer is one item from a list of strings
score_fields | Names the rubric fields that return an integer and an expected score

Example: Pass a short note saying returns are accepted for 30 days and the item was bought 12 days ago, with a true or false field named eligible. You get back the chosen output and the probability the model gives each answer for that field.

For newcomers and experts: If you are new, think of it as a judge that picks from options you define. Try the return-window sample first with a boolean field, then add a choice list once that works.

For experts, it scores the allowed answer tokens directly and applies a softmax over the candidate logits once, with a default temperature of 1.0. Fields are handled one at a time. Each field takes at most 255 choices, and a prompt over 8,192 tokens is rejected instead of cut. The card says this checkpoint has no separate temperature calibration, so do not reuse the older model's value.

Define a context and a schema of fields
Call score on the loaded model
The adapter scores each allowed answer
A softmax gives probabilities per option
You read the output and the probabilities

Good to know: The points below cover limits and costs.

• Hardware: a CUDA GPU with BF16 support is expected.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. A Hugging Face token is needed only for a private repository, and you need a capable GPU.
• Data it reaches: it reads the context you pass in. The base Qwen3.5-9B weights come from Hugging Face and are not copied into this repository.
• Not for: free-form answers or generated reasoning, prompts over 8,192 tokens, or more than 255 choices in one field.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 1,027
• Likes: 254
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-9B
• Tags: peft, safetensors, lora, qwen3.5, text-classification, structured-prediction, evidence-grounding

Links

• Model: https://huggingface.co/bespokelabs/Bespoke-Nimble-9B
• Files: https://huggingface.co/bespokelabs/Bespoke-Nimble-9B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download bespokelabs/Bespoke-Nimble-9B

Then follow the usage in its model card. Nothing in this kit runs by itself.