Perplexity AI has launched Sonar, a new AI search model built on Meta's Llama 3.3 70B architecture and powered by Cerebras Inference, claiming a significant leap in speed and performance. This move positions Perplexity to challenge established players like Google and Bing in the competitive search engine market.
Sonar boasts an impressive processing speed of 1200 tokens per second, enabling near-instantaneous responses to user queries. Perplexity emphasizes that Sonar has been optimized for both factuality and readability, two key factors impacting user satisfaction. Internal evaluations suggest that Sonar outperforms OpenAI's GPT-4o mini and Anthropic's Claude 3.5 Haiku, and achieves performance parity with larger models like GPT-4o and Claude 3.5 Sonnet, while being significantly faster than Google's Gemini 2.0 Flash.
Currently available to Perplexity Pro users, Sonar can be set as the default model through the platform's settings. This release follows the recent launch of the Sonar API, offered in both Sonar and Sonar Pro versions, touted as the most affordable API on the market. The Pro version offers enhanced capabilities for complex, multi-step tasks requiring deeper understanding and context retention, along with more comprehensive citations.
This partnership with Cerebras Systems highlights the growing trend of AI companies leveraging specialized hardware for competitive advantage. The use of Cerebras's Wafer Scale Engine is a key factor in Sonar's speed and efficiency. Perplexity's focus on developer accessibility through the API further broadens Sonar's potential impact, opening doors for integration into a variety of applications. This strategic approach, combined with competitive pricing, positions Sonar as a compelling alternative for developers seeking powerful and efficient AI-driven search capabilities. While currently limited to Pro users, the wider availability of Sonar could significantly disrupt the search engine landscape.