Sonar (Perplexity)

Sonar is Perplexity's in-house language model designed specifically for the company's default search experience, where the priority is producing clean, factual answers quickly rather than open-ended generation. According to Perplexity's own announcement, Sonar was built on top of Meta's Llama 3.3 70B and then further trained by Perplexity to improve factuality and readability in answer-style responses. That lineage matters for practitioners: it is a fine-tuned derivative of a well-known open base model, but Perplexity positions it as its own proprietary system optimized for grounded, search-driven question answering rather than as a general chat model.

In Perplexity's online A/B testing, Sonar was reported to substantially outperform smaller competitors such as GPT-4o mini and Claude 3.5 Haiku on user satisfaction, and to closely match or exceed larger frontier models like GPT-4o and Claude 3.5 Sonnet for the same metric. To reach that experience, Perplexity serves the model through Cerebras inference infrastructure, achieving a reported decoding speed of roughly 1200 tokens per second, which enables near-instant answer generation in the search interface. For developers, Sonar is also offered as an API aimed at embedding Perplexity-style AI search into third-party applications, making it a practical fit when low-latency, citation-friendly, answer-style outputs matter more than maximum reasoning depth.