Perplexity Sonar Reasoning

O Perplexity AI

Osnovano 2022 · San Francisco, California, USA

Perplexity AI was founded in August 2022 by Aravind Srinivas (CEO, former OpenAI and DeepMind researcher), Denis Yarats (former Quora), Johnny Ho (former Quora) and Andy Konwinski (Databricks co-founder). Headquartered in San Francisco, Perplexity built one of the first AI-native search products — an answer engine combining live web retrieval with LLM synthesis and inline citations. The company has raised over $500M from investors including IVP, Nvidia, NEA, Jeff Bezos and SoftBank, with a 2024 valuation above $9B. Perplexity offers its Sonar API for developers to access search-grounded LLM responses. Sonar Reasoning, released January 2025, is built on the open-source DeepSeek-R1 reasoning model wrapped in Perplexity's live web-search retrieval stack with the reasoning trace exposed in the API response.

Arhitektura

Reasoning Mixture-of-Experts Transformer with live web retrieval

Sonar Reasoning is Perplexity's hosted deployment of DeepSeek-R1 (671B Sparse Mixture-of-Experts, 37B active per token, DeepSeekMoE architecture with Multi-head Latent Attention) wrapped in Perplexity's retrieval pipeline. The base model uses 256 fine-grained experts plus shared experts, top-8 routing, 128K context, and was trained by DeepSeek with the GRPO reinforcement-learning recipe described in the R1 technical report. Perplexity hosts the model on US-based infrastructure, applies its own safety post-training on the Perplexity domain, and exposes the reasoning chain-of-thought as a separate field in the API response alongside the final answer. Each request triggers a live web-search retrieval that injects ranked passages into the model context with source URLs returned alongside the answer. Released January 2025 via the Sonar API.

Parametri

671B total, 37B active per token (DeepSeek-R1 base)

Kontekst

127.000 tokena

Mogućnosti

• Combines live web search with explicit reasoning traces
• Returns ranked citations with source URLs alongside answers
• Strong on math, logic and research-style questions
• Reasoning chain-of-thought exposed in the API
• Knowledge always current via live search — bypasses model cutoff limits
• Cheaper than OpenAI o-series for comparable reasoning quality at release
• Hosted on Perplexity's US infrastructure
• Best for: research-grade Q&A, market intelligence, citation-required workflows, fact-grounded reasoning.

Obuka i licenca

Inherits DeepSeek-R1's training pipeline (DeepSeek-V3-Base ~14.8T tokens of multilingual web, code and math, followed by the R1 reasoning RL stage with GRPO). Perplexity adds retrieval-grounded supervised fine-tuning on its domain. Base model knowledge cutoff approximately July 2024 — but live search overrides for current questions.

Licenca: Perplexity API Terms of Service. The base DeepSeek-R1 weights are MIT-licensed and can be self-hosted separately, but the live-search and reasoning-citation pipeline are Perplexity-specific and hosted-only.

Testiranje sigurnosti: Perplexity applies safety post-training and operates the model under its hosted API safety policies. The base DeepSeek-R1 has its own published model card; Perplexity's hosting modifies refusal behaviour for the Sonar deployment.