basal-omni-1.0-ultra

basal-omni-1.0-ultra

Answers typed questions about a message, document or agent trace with a calibrated probability for each allowed answer, in one pass without generating text. An experimental 31B model for Polish and English decisions.

Made by Remek 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: basal-omni-1.0-ultra is an experimental 31B model built on Google Gemma 4, the largest in the basal-omni line. It does not write answers. It reads a state, such as a message, a document, retrieved passages or an agent trace, and answers a typed question about it. The reply is a probability for each allowed option.

Who it suits: Teams who need fast, calibrated decisions on Polish or English text, for example routing, labeling or judging content.

Industry | Who uses it | How they use it | Example
Software | Developers | Add a decision step to a pipeline | Pick the right tool for an agent request
Operations | Process owners | Label or route documents automatically | Classify an incoming contract
Research | ML researchers | Study calibrated decision models | Compare it with the older basal-1.5-max

If you need a model that writes text, you can skip it, since it only picks among options you give.

Where to use it:

• The basal inference engine, which serves the typed questions over its API
• Agent pipelines that choose among OpenAI or MCP tool definitions
• Decision tasks on Polish and English text, plus images in the same request

Use case | Who | What they get back
Tool choice | Agent builders | The tool to call, not its arguments
Yes or no checks | Developers | A probability for yes and for no
Scoring and choosing | Analysts | A probability for each option, up to 255 of them

How to set it up:

• Python with pip, to fetch the files
• The basal engine, which the card names as the way to serve the model
• A large GPU, since the card's results were served on one H200

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Serve them with the basal engine, following its instructions on GitHub.

pip install -U huggingface_hub
hf download Remek/basal-omni-1.0-ultra

How to use it: You send the engine a state and a typed question with the allowed answers. It returns a probability for each one. The question types the card names are in the table. Images go in the same request when the server runs with the media option on.

Tool | What it does
choice | Picks among the options you give
noul | Answers yes or no
score | Rates the state on a scale
tool | Picks a tool from OpenAI or MCP tool definitions

Example: You pass a customer message as the state and ask a yes or no question about whether it is a complaint. The model returns a probability for yes and for no, so you can act automatically when it is sure and send it to a person when it is not.

For newcomers and experts: If you are new, think of it as a fast classifier that also tells you how sure it is. It cannot answer outside the options you give it.

For experts, it is inspired by System 1 style decision models. It handles images zero-shot through Gemma 4's own image encoder, and takes no audio.

You send a state and a typed question
The model reads them in one forward pass
It returns a probability per allowed option
Your code acts or defers to a person

Good to know: The card's Werdykt v1 benchmark compares three models on hidden Polish and English tasks.

Title: Werdykt v1 macro score from the card
basal-omni-1.0-ultra: 0.879 score
basal-omni-1.0-max: 0.823 score
basal-1.5-max: 0.773 score

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need a large GPU and the basal engine.
• Data it reaches: downloading reaches Hugging Face. The model only reads the state you send it.
• Not for: writing text, audio input, or exact rules. The card says it is experimental, and its rules category scored 0.550, so compute exact thresholds and dates in code.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 31.3B
• Weights: about 62.5 GB in BF16; about 15.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,831
• Likes: 43
• License: apache-2.0
• Languages: pl, en
• Based on: google/gemma-4-31B-it
• Quantized versions: 3
• Tags: transformers, safetensors, gemma4, image-text-to-text, typed-decisions, decision-making, classification, calibration, polish, experimental, decision-model, text-generation

Links

• Model: https://huggingface.co/Remek/basal-omni-1.0-ultra
• Files: https://huggingface.co/Remek/basal-omni-1.0-ultra/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Remek/basal-omni-1.0-ultra

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