Nimble-v4-max

Nimble-v4-max

Gives a probability for every answer option when you pass a context and typed questions such as choices, yes or no, or ordered scores. A fine-tune of Qwen3.8-27B for builders of classifiers and decision steps.

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: Nimble-v4-max is a decision model from Bespoke Labs, fine-tuned from Qwen3.8-27B. You give it a context and typed questions, such as picking one of several options, answering yes or no, or rating on an ordered scale. It gives back a probability for every option. This page covers version 4.1, which replaces 4.0 and accepts the same requests.

Who it suits: It suits developers who build classifiers, routers or scoring steps and want probabilities instead of free text. Teams that need a calibrated confidence for each answer option can use it too.

Industry | Who uses it | How they use it | Example
Customer support | Developers of chat systems | Sort a customer message into an intent | Each possible intent gets its own probability
Research | Machine learning researchers | Score answers to multiple-choice questions | Every option comes back with a probability
Product teams | Builders of review or triage steps | Ask a yes or no question about a text | The answer comes with a probability for yes and for no

You can skip it if you want free-form text, since loading it as a plain chat model does not reproduce its answers.

Where to use it:

• Python with the transformers library
• One GPU with 80 GB of memory or more for the BF16 weights
• Windows, macOS or Linux, to download the files from Hugging Face
• Your own pipeline, through the inference code Bespoke Labs publishes

Use case | Who | What they get back
Intent classification | Developers | A probability for each intent label
Select-all and yes or no questions | Analysts | A probability for every option
Benchmark-style reasoning checks | Researchers | A ranked set of answer options

How to set it up: You need Python, a large GPU and the published inference code.

• Python with the huggingface_hub package
• A GPU with 80 GB or more, since the BF16 weights need about 55 GB
• The inference code from the results dataset of the model, in its code/nimble folder

1. Install the Hugging Face tools with Python.
2. Download the model files with the second command below.
3. Fetch the inference code from the results dataset of the model.
4. Send your requests through that code, not through a plain chat setup.

pip install -U huggingface_hub
hf download bespokelabs/Nimble-v4-max

How to use it: You pass a context and your typed questions to the inference code. It builds the prompts, applies the attention pattern used in training, and scores a single token for every option. By default it applies a softmax at temperature 1. To calibrate the probabilities, divide the logits by the model's default temperature before the softmax. The top-ranked option stays the same whatever the temperature.

Tool | What it does
code/nimble | Inference code in the results dataset that builds the prompts and scores the options
schema_config.json | Pins the prompts by hash
benchmark-summary.json | Holds scores for all 43 benchmarks in the results dataset

Example: You pass a customer message and a question asking which of several intents it matches. The model returns a probability for each intent, and the highest one is its answer. Divide the logits by the default temperature before the softmax if you want calibrated probabilities.

For newcomers and experts: Think of it as a multiple-choice grader, not a chatbot. You do not read text from it. You read a probability next to each option. A first small try is a single yes or no question run through the published inference code.

It is a fine-tune of Qwen3.8-27B that scores one token per option instead of writing text. The inference code builds the prompts, which the schema config pins by hash, and uses the attention pattern from training. Plain chat loading or text generation will not match its answers. Temperature rescales the logits and never changes the ranking.

Build prompts from context and questions
Apply the training attention pattern
Score one token per option
Softmax over the option logits
Divide by temperature to calibrate

Good to know: Version 4.1 beats version 4.0 on the public Decision Index 0.3 tests, with the same requests and code.

Title: Decision Index 0.3 public score
v4.0: 68.46
v4.1: 72.74

Title: Selected benchmark scores
Benchmark | v4.0 | v4.1
MMLU-Pro | 71.0 | 74.1
GSM8K | 75.9 | 83.6
BANKING77 | 79.9 | 93.2
CLINC150 + OOS | 91.6 | 96.2

To stay on version 4.0, pin the revision:

revision="v4.0"

• Size: 27.4B parameters, about 54.7 GB in BF16 and about 13.7 GB at 4-bit, which is an estimate
• License: CC BY-NC 4.0, so no commercial use, and the Qwen3.8-27B base model has its own Apache 2.0 license
• Cost to run: none named, but it needs one GPU with 80 GB or more
• Data it reaches: none named beyond the downloaded files and the context you pass in
• Not for: commercial use, plain chat or free text generation

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 366
• Likes: 7
• License: cc-by-nc-4.0
• Based on: Qwen/Qwen3.8-27B
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

• Model: https://huggingface.co/bespokelabs/Nimble-v4-max
• Files: https://huggingface.co/bespokelabs/Nimble-v4-max/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download bespokelabs/Nimble-v4-max

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