Remek/basal-omni-1.0-max
Remek/basal-omni-1.0-max is an experimental 12-billion parameter model from basal's omni line, built on Google Gemma 4 with a 32768-token context length. It specializes in making calibrated, typed decisions (e.g., choice, yes/no, score) from text, images, and audio in a single forward pass, providing probability scores for each option without generating text. This model is optimized for structured decision-making tasks, offering up to 255 options per question and supporting zero-shot image and audio processing via Gemma 4's encoders. It excels in classification, verification, compression, and extraction tasks, particularly for Polish and English documents.
Loading preview...
basal-omni-1.0-max: A Multimodal Decision-Making Model
basal-omni-1.0-max (12B) is an experimental model from Remek's basal-omni line, built upon the Google Gemma 4 base. Unlike traditional LLMs that generate text, this model focuses on making calibrated, typed decisions in a single forward pass, providing a probability for each allowed answer. It processes text, images, and audio (zero-shot via Gemma 4's encoders) to answer structured questions, making it suitable for automated decision systems.
Key Capabilities
- Typed Decision-Making: Answers questions with calibrated probabilities for predefined options (e.g.,
choice,noul,score,multi,act,tool), supporting up to 255 options per question. - Multimodal Input: Processes text, images, and audio within the same request, leveraging Gemma 4's encoders for zero-shot media understanding.
- Structured Tasks: Supports
classify,verify,compress, andextracttasks, expanding them into internal decision processes. - Performance: Achieves strong results on decision benchmarks like Werdykt v1 (0.823 macro score) and various Polish and English decision tasks, outperforming basal-1.5-max.
- Efficiency: Provides calibrated confidence scores, allowing for automated decision-making above a certain threshold and routing uncertain cases to human review.
Good for
- Automated Classification & Routing: Ideal for categorizing customer inquiries, documents, or other inputs into predefined categories with high confidence.
- Information Verification: Useful for verifying statements against provided text or media, determining if information is supported, contradicted, or not stated.
- Structured Data Extraction: Efficiently extracts specific fields from documents (e.g., invoices) into a JSON schema.
- Multimodal Decision Systems: Applications requiring decisions based on a combination of text, images (e.g., scanned documents), and audio (e.g., spoken requests).
- High-Throughput Decision Engines: Designed for rapid, calibrated decision-making in scenarios where text generation is not required.