Sonar Reasoning

Sonar Reasoning is a large language model developed by Perplexity AI, designed to provide high-accuracy answers by combining real-time web search with advanced chain-of-thought (CoT) reasoning. The model is part of the Sonar family and is specifically optimized for complex, multi-step queries that require logical deduction and verified information retrieval.\n\nThe model architecture integrates the DeepSeek-R1 reasoning framework, allowing it to generate internal reasoning tokens before producing a final response. This process enables the model to self-correct, plan its search strategy, and synthesize information from multiple web sources with precision. By leveraging Perplexity's proprietary search index, Sonar Reasoning grounds its logic in up-to-date data, which helps in maintaining factual accuracy during long-form research tasks.\n\n## Technical Capabilities\nSonar Reasoning supports a context window of 128,000 tokens, making it capable of processing extensive search results and maintaining large conversational histories. It is frequently applied to analytical tasks such as technical research, coding assistance, and data synthesis. The model outputs cited responses, linking directly to the primary sources retrieved during its search phase, providing transparency and verifiable evidence for its conclusions.