Perplexity: Sonar

Perplexity: Sonar is a text model from Perplexity, released on 27 January 2025. Perplexity describes it as lightweight, fast, affordable and simple to use, with citations and the ability to customize sources. It is aimed at companies that want to integrate lightweight question and answer features into their own products rather than build a general purpose assistant from scratch.

The model accepts text and images as input and returns text. It has a context window of 127,072 tokens and a maximum output of 114,364 tokens, so it can hold long inputs and produce long answers in a single call. The combination of image input with text output means it can read a screenshot, diagram or photo alongside a written question and answer in prose. Because the provider description highlights citations and customizable sources, answers can be grounded in sources you choose, which matters when you need to show where an answer came from.

Those traits point to a fairly specific set of uses. Question and answer features inside a product are the obvious fit: support widgets, documentation helpers, internal knowledge lookups and similar flows where a user asks something and expects a short, sourced reply. Image input extends this to cases where the question is about something visual, such as asking about the contents of a screenshot or a photographed document. The large output ceiling suits tasks that need a long written response, such as summarizing several documents or drafting a detailed explanation, while the large context window lets you pass in substantial background material without splitting it up.

You call the model through api14's OpenAI-compatible API using the model id "perplexity/sonar". The supported request parameters are frequency_penalty, max_tokens, presence_penalty, temperature, top_k, top_p and web_search_options. The first six are the usual sampling and length controls, so you can tune how deterministic or varied the output is and cap how many tokens come back. web_search_options is the parameter to reach for when you want the model to draw on web search and is the main lever for the source customization the provider mentions.

There is no information here about rate limits, latency figures or accuracy, so those need testing against your own workload. If your task is pure text generation with no need for citations or source control, a plain text model may be simpler. Sonar makes most sense when grounded, cited answers and optional image input are central to what you are building.