Remek/basal-omni-1.0-mini
Remek/basal-omni-1.0-mini is an experimental 7.9 billion parameter model based on Google's Gemma 4 E4B architecture, designed for typed decision-making. It processes text, images, and audio (zero-shot for media) to provide calibrated probabilities for up to 255 predefined options in a single forward pass, without generating text. This model excels at tasks requiring precise, probabilistic answers to structured questions, offering improved performance over its predecessor, basal-1.5, particularly in English and Polish decision benchmarks.
Loading preview...
basal-omni-1.0-mini: Experimental Multimodal Decision Model
basal-omni-1.0-mini is an experimental 7.9 billion parameter model from Remek, built on Google's Gemma 4 E4B architecture. It is the smaller counterpart to basal-omni-1.0-max (12B) and represents a novel approach to decision-making, inspired by System 1 intuitive models.
Unlike traditional LLMs that generate text, this model reads a state (text, images, or audio) and answers typed questions by providing a calibrated probability for each allowed answer in a single forward pass. This ensures answers never fall outside the given options and quantifies the model's certainty.
Key Capabilities
- Typed Decision-Making: Answers structured questions with probabilistic scores (e.g.,
choice,yes/no,score,multi,act,tool). - Multimodal Input (Experimental): Processes images and audio alongside text in a zero-shot manner, utilizing Gemma 4's encoders without prior media training. This feature is currently experimental, with media training planned for future releases.
- High Option Capacity: Supports up to 255 options per question, a significant increase from basal-1.5's 10 options.
- Calibrated Confidence: Provides a confidence score for each decision, allowing for automated acceptance of high-confidence answers and human review for others.
- Improved Performance: Outperforms basal-1.5 (4.5B) on several benchmarks, including Werdykt v1 (macro 0.769 vs 0.721) and English external panel tasks (0.740 vs 0.688).
Good For
- Automated Triage and Routing: Quickly categorizing customer inquiries or documents based on predefined criteria.
- Structured Data Extraction: Extracting specific, probabilistic answers from complex documents or messages.
- Agentic Workflows: Providing reliable, probabilistic decisions within agent systems, especially for tool selection or action determination.
- Applications requiring high certainty: When a precise, quantifiable confidence in a decision is critical, rather than a free-form text generation.
- Processing mixed media inputs: For use cases where decisions need to be made based on a combination of text, images, and audio, even in its current experimental zero-shot media capability.