Smaug-Decisions

Smaug-Decisions

Answers a typed question about audio, an image or text with a probability for each option, in one pass and with no generated text. A 4B decision model fine-tuned from Gemma 4 E4B-it.

Made by abacusai 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: Smaug-Decisions is a 4B decision model from Abacus.AI that hears, sees and reads. Abacus.AI built it by further training Google's Gemma 4 E4B-it. You give it audio, an image or text plus a question with a few options, and it returns a calibrated probability for each option in a single forward pass, with no generated text.

Who it suits: It suits developers who need a quick, thresholdable decision from audio, images or text, such as detecting whether a speaker has finished talking.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Voice agents | Developers | Decide from audio whether a speaker is done talking | A yes or no probability for a call recording
Audio analytics | Product teams | Ask typed questions about audio clips and act on the probabilities | Setting a threshold on one option's probability
Mobile apps | App developers | Run decisions offline on a phone through the LiteRT version | Answering a text question on an iPhone without a network
Research | ML researchers | Compare decision models by input type | Reading the Earshot and Jev Decision Index scores

People who need generated text or audio longer than 30 seconds can skip it.

Where to use it:

• Transformers on an NVIDIA GPU in bfloat16, on Apple Silicon through MPS, or on a CPU
• Phones, through the separate Smaug-Decisions-LiteRT version
• A faster tier: the card links a 0.7B audio-only sibling
• Windows, macOS or Linux: download the files with Python

Title: Jobs it does
Use case | Who | What they get back
Audio decisions | Voice developers | A probability for each option, from clips up to 30 seconds
Image questions | App teams | A probability for each option about an image
Text questions | Developers | A probability for each option about a piece of text

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• The huggingface_hub and Transformers libraries
• Storage for about 17.2 GB of BF16 weights, or an estimated 4.3 GB at 4-bit

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load the model with the helper script that ships in the repository, as the usage code shows.

pip install -U huggingface_hub
hf download abacusai/Smaug-Decisions

How to use it: Load the model with the helper script, then call it with an input and a question. It returns a probability for each option.

import sys
from huggingface_hub import snapshot_download
sys.path.insert(0, snapshot_download("abacusai/Smaug-Decisions", allow_patterns=["smaug_decide.py"]))
from smaug_decide import SmaugDecide                                   # ships with this repo

model = SmaugDecide.from_pretrained("abacusai/Smaug-Decisions")
model("call.wav", "Has the speaker finished their turn?")               # -> {'Yes': 0.93, 'No': 0.07}

The input can be audio up to 30 seconds at 16 kHz mono, an image, or text. The question carries 2 to 20 options. One answer uses a softmax, and a select-all question uses a sigmoid.

Example: You pass a call recording and ask whether the speaker is done talking. The model returns a probability for each option, as in the card's sample of 0.93 for yes and 0.07 for no.

For newcomers and experts: If you are new to this, think of it as a judge that answers a multiple-choice question about a sound, picture or text, and tells you how sure it is. Try one short audio clip and a yes or no question first.

If you work in the field, the model reads out a letter logit for each option instead of decoding text. Counted by compute it is a 4B model, while 8.6B parameters are stored, because 3.5B of them sit in embedding tables that are only looked up. The card says its probabilities are better calibrated than the frontier APIs it compares against, with an expected calibration error of 0.069.

Pass audio, an image or text
Add a question with 2 to 20 options
One forward pass reads out the options
A probability per option comes back

Good to know: Abacus.AI measures audio skill on its own Earshot benchmark of 6,671 items. On the public Jev Decision Index the model reaches 57.9 for images and 56.6 for text.

Title: Earshot audio skill
Smaug-Decisions: 60.1 points
Gemini 4 Argon: 59.4 points
Gemini 3.8 Flash: 50.5 points

• License: apache-2.0
• Cost to run: none named
• Data it reaches: the audio, images and text you pass in
• Not for: open-ended text generation, since it answers with option probabilities, or audio longer than 30 seconds

On Hugging Face, 13 November 2026

• Task: any to any
• Library: transformers
• Parameters: 8.6B
• Weights: about 17.2 GB in BF16; about 4.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 110
• Likes: 4
• License: apache-2.0
• Languages: en
• Based on: google/gemma-4-E4B-it
• Quantized versions: 1
• Tags: transformers, safetensors, gemma4, image-text-to-text, decisions, typed-decisions, audio, vision, calibration, on-device, smaug, abacusai

Links

• Model: https://huggingface.co/abacusai/Smaug-Decisions
• Files: https://huggingface.co/abacusai/Smaug-Decisions/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download abacusai/Smaug-Decisions

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