Cerebras and Perplexity Team Up to Build a New AI Search Model: Sonar 240

Cerebras has recently partnered with Perplexity AI to launch Sonar, an extremely fast AI search model. The Sonar model is built on Cerebras's dedicated AI chip and is capable of running at 1200 tokens per second, making it one of the fastest AI search models on the market.

The Sonar model is fundamentally based on Llama 3.3 70B and has been further optimized to enhance performance for search applications. In this release, Perplexity emphasized that Sonar is not only extremely fast but also provides higher-quality answers, generating responses almost instantly.

Another highlight of Sonar is its contextual understanding capability, which allows it to better handle multi-turn conversations, remember user context, and thereby deliver more personalized responses—such as handling follow-up questions—enabling users to obtain relevant results when making complex queries.

In addition, Sonar offers multimodal response functionality, supporting responses that combine images and text. For example, when we query information about a certain product, Sonar can not only provide a textual description but also display relevant images or charts.

In terms of security and privacy, Perplexity AI places special emphasis on protecting user data, with all query data being processed through anonymization.

Cerebras's AI inference is considered a key advantage of the Sonar model, allowing it to surpass traditional search engines across many performance metrics. This collaboration aims to challenge the current search market, including major competitors like Google, and is expected to capture a market share worth hundreds of billions of dollars.

Perplexity has also announced that it will provide a Sonar API service, empowering developers and enterprises to build auto-generative search engines.

According to recent information, the Sonar model significantly outperforms other well-known AI models in user satisfaction and accuracy metrics, such as GPT-4o mini and Claude 3.5 Haiku. This innovative application achieved a factuality score of 85.1, compared to 83.9 for GPT-4o and 75.8 for Claude 3.5 Sonnet, demonstrating Sonar's outstanding performance in search accuracy.