joey-1

joey-1

Returns calibrated probabilities for yes-or-no, multiple-choice and scored questions about a text or an image, from one 2B model. For teams that route, triage or classify requests automatically.

Made by baselabs 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: joey-1 is a decision model with 2B parameters, trained by Base Labs at Baseten from Qwen3.5-2B. You give it a situation described in text, optionally with images, plus a set of questions. It replies with a probability for each answer option and handles all the questions in one pass.

Who it suits: It suits developers who need an automatic judgment step, such as sending a message to the right team or deciding how urgent it is, with probabilities they can act on.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Route incoming messages to the right team | Deciding whether a message is a refund request
Back-office automation | Operations teams | Score how urgent an item is | Ranking a complaint from can wait to blocking
Software agents | Agent developers | Add a quick decision step inside a workflow | Choosing between billing, technical and fraud handling
Classification work | Data teams | Label items with calibrated confidence | Using the probabilities to flag uncertain cases

You can skip it if you need free-form text generation, since it answers with option probabilities.

Where to use it:

• Baseten Model APIs, which host it and accept text and image input
• Your own server with one CUDA GPU and the serving runtime the card pins
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Yes or no check | Developers | The probability of yes for a question
Pick a team | Support teams | A chosen option and a probability for each named one
Rate urgency | Operations | A level on an ordered scale, plus probabilities

How to set it up: Self-hosting needs one CUDA GPU and a specific PyTorch container image. The hosted route needs no setup on your side.

• One CUDA GPU
• The PyTorch 2.8.0 container image for CUDA 12.8 with cuDNN 9, in its devel flavor
• Python and huggingface_hub to download the files

1. Download the model files into a joey-1 folder.
2. Run the setup script, which installs the pinned serving runtime.
3. Use the Python engine in the next section to ask your questions.

hf download baselabs/joey-1 --local-dir joey-1 && bash joey-1/setup.sh

How to use it: Create a decision engine, define your questions with a type for each, and call decide with the situation text. The script below asks three questions about one customer message and prints the answers.

from qev.engine import DecisionEngine
from qev.schema import SystemOneRequest

engine = DecisionEngine("baselabs/joey-1", device="cuda:0")

questions = {
    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {"billing": "Charges, refunds, invoices",
                     "technical": "App or site faults",
                     "fraud": "Suspected unauthorised use"},
    },
    "urgency": {
        "type": "score",
        "instructions": "How urgent is this?",
        "criteria": ["Can wait", "Today", "Blocking the customer now"],
    },
}
resp, _ = engine.decide(SystemOneRequest(
    state="I was charged twice this month, please refund one of them.", questions=questions))
print(resp.answers)

Tool | What it does
noul | Asks a yes or no question and returns the chance of yes
choice | Picks among named options, with a probability for each
score | Rates on ordered levels, lowest first, with probabilities

Example: You send a customer note saying a card was charged two times and one payment should come back, along with the refund, team and urgency questions. The answers hold a likely yes for the refund question and a probability for each team and urgency level.

For newcomers and experts: If you are new to this, think of a form with a few questions about a message, answered by a model that also says how sure it is. A first small try is the script above with your own sentence.

For people who know the field, it is a full fine-tune of the language model with the vision encoder frozen, trained with cross-entropy plus a Brier calibration term. One shared-prefix prefill serves all the questions, and the maximum sequence length is 16,384 tokens.

Title: Decision Index on the Jev 0.4 public split
elephant-1: 72.24
joey-1: 63.30
Cloudflare Clef: 61.71
Jev 1.14: 60.70
Perplexity Decider Flash v1.2: 60.06

You send a situation and your questions
One shared prefill reads the text and images
The model scores every option of every question
You get a probability for each option

Good to know: Inputs that exceed the lengths it was trained on get rejected instead of trimmed, and the benchmark scores in the card come from its authors.

• License: MIT
• Cost to run: the card lists Baseten Model APIs at 0.01 dollars per million input tokens, or your own CUDA GPU when self-hosted
• Data it reaches: the text and images you send, with zero data retention on by default on the hosted API
• Not for: open-ended text generation, or inputs longer than the trained limits

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 2.7B
• Weights: about 5.4 GB in BF16; about 1.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,539
• Likes: 47
• License: mit
• Languages: en
• Based on: Qwen/Qwen3.5-2B
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, decision-model, classification, calibration, routing, guardrails, vision, multimodal, baseten

Links

• Model: https://huggingface.co/baselabs/joey-1
• Files: https://huggingface.co/baselabs/joey-1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download baselabs/joey-1

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