Sonar

Sonar is Perplexity's in-house language model, designed from the ground up for the answer-generation layer behind Perplexity's default search experience. Rather than a general-purpose chat model, its purpose is narrow and deliberate: produce readable, factual responses to natural-language questions in near real time. Perplexity describes the model as optimized for answer quality and user experience, with a clear emphasis on factual reliability and clear prose when responding to live search queries.

According to Perplexity's own launch announcement, Sonar is built on top of Meta's Llama 3.3 70B base model and has been further trained to strengthen answer factuality and readability for search-style use. The same source credits Sonar with a decoding throughput of roughly 1200 tokens per second on Cerebras inference hardware, and reports that in Perplexity's online A/B testing it surpassed smaller peers such as GPT-4o mini and Claude 3.5 Haiku on user satisfaction while approaching the performance of larger frontier models like GPT-4o and Claude 3.5 Sonnet. That combination of a strong open-base lineage, search-targeted post-training, and unusually fast inference points to a practical fit for latency-sensitive retrieval-augmented applications where concise, citable answers matter more than open-ended reasoning.