sui-1: Grounded and Verifiable Long-Form Summarization
sui-1 (Summarization with Unique Identifiers) by ellamind is a specialized 24 billion parameter model designed for high-quality summarization of extremely long documents, capable of processing up to 2 million tokens through an iterative approach. Its core differentiator is a built-in source grounding mechanism: every claim in the summary is linked via XML tags to its exact source sentence in the original text, enabling full traceability and significantly mitigating hallucination.
Key Capabilities
- Verifiable Outputs: Each summary claim is traceable to its source sentence, ensuring factual accuracy and reducing hallucination.
- Very Long Document Processing: Natively handles up to 128k tokens, with an iterative two-step approach extending support to documents up to 2 million tokens.
- Efficient Deployment: The FP8 quantized variant can run on a single A100 40GB or A6000 48GB GPU, making it accessible for more modest hardware.
- Multilingual Support: Fine-tuned for English, German, Spanish, French, and Italian, and inherits support for 20+ additional languages from its Mistral Small 3.2 base.
- High-Quality Training: Trained on over 22,000 examples generated with a sophisticated chain-of-thought reasoning and multi-stage verification pipeline.
Good for
- Enterprise-grade summarization: Ideal for applications requiring high factual accuracy and auditability, such as legal, financial, or research document analysis.
- Processing extensive reports: Effectively summarizes very long texts like parliamentary documents, academic papers, or large web compilations.
- Resource-constrained environments: The FP8 variant and iterative processing allow deployment on GPUs with limited VRAM, making advanced summarization more accessible.
- Multilingual content analysis: Provides high-quality summarization across several major European languages.