Sonar is a suite of language models developed by Perplexity AI, specifically engineered for conversational search and retrieval-augmented generation (RAG). By fine-tuning base architectures like Meta's Llama on proprietary search data, Perplexity optimized Sonar to produce factual, grounded responses with integrated citations. The models are designed to minimize hallucinations while maintaining high throughput for real-time web-connected queries.
The family consists of multiple tiers tailored for different complexity levels. Sonar serves as the efficient, low-latency option for standard queries, while Sonar Pro provides more advanced reasoning capabilities for complex research tasks. In late 2024, the line-up was further expanded with Sonar Reasoning, which incorporates chain-of-thought logic to improve performance on sophisticated logical and mathematical problems. Unlike standard general-purpose models, the Sonar series is purposefully constrained to prioritize accuracy and verification against external sources, making it a specialized tool for search-centric applications where source transparency is a requirement.