Perplexity AI answer engine combining a proprietary web index with LLMs to deliver cited, conversational search.
Perplexity is the highest-profile AI-native search company: it runs its own crawler and 200B+ URL index on Vespa, post-trains its own Sonar models on open-weights Llama bases, and serves them on a largely self-built GPU inference stack (Rust/CUDA engine, open-sourced RDMA kernels) spread across AWS and other clouds. At the same time it is pragmatically multi-vendor at the model layer, reselling GPT, Claude, Gemini and Grok in its Pro model picker, and it publishes an unusual amount about its infrastructure via its research blog, vendor case studies, and job postings.
perplexity.aiSources last verified Jul 26, 2026 · 2 historical entries retained
In-house search model family powering default search.
Receipts · 1
“Built on the robust foundation of Llama 3.3 70B, Sonar represents a significant advancement in answer quality, factuality, and readability” — cerebras.ai, Feb 11, 2025
Source last verified Jul 26, 2026
Receipts · 2
“different variations of the popular open source Llama 3.1 models like 8B, 70B, and 405B” — developer.nvidia.com, Dec 5, 2024
“Built on the robust foundation of Llama 3.3 70B” — cerebras.ai, Feb 11, 2025
Source last verified Jul 26, 2026
Receipts · 1
“Perplexity uses Amazon Bedrock to serve Claude models on both the free and paid versions of its search engine” — claude.com, Jul 2024
Source last verified Jul 26, 2026
Receipts · 1
“GPT-5 is now available on Perplexity and Comet for Max and Pro subscribers.” — x.com, Aug 7, 2025
Source last verified Jul 26, 2026
Post-trained open-weights reasoning model with censorship removed.
Receipts · 1
“We are not able to make use of R1's powerful reasoning capabilities without first mitigating its bias and censorship” — eweek.com, Feb 20, 2025
Source last verified Jul 26, 2026
Custom Rust and CUDA engine serving every query across a multi-cloud GPU fleet.
Receipts · 2
“We build and run the inference engine behind every Perplexity query and deploy dozens of model architectures at scale” — jobs.ashbyhq.com, Apr 13, 2026
“a large GPU fleet spread across several cloud providers” — jobs.ashbyhq.com, Jul 16, 2026
Source last verified Jul 26, 2026
TensorRT-LLM
by NVIDIA Served LLMs on GPU pods before the in-house engine took over.
Receipts · 2
“Triton Inference Server is a critical component of Perplexity's deployment architecture.” — developer.nvidia.com, Dec 5, 2024
“Our stack is Rust, Python, CUDA, and CuTe DSL” — jobs.ashbyhq.com, Apr 13, 2026
Source last verified Jul 26, 2026
Cerebras Inference
by Cerebras Wafer-scale inference serving the Sonar model for near-instant answers.
Receipts · 1
“1,200 tokens per second, delivering near-instant answer generation” — cerebras.ai, Feb 11, 2025
Source last verified Jul 26, 2026
Open-sourced RDMA library powering disaggregated serving of large MoE models.
Receipts · 2
“KvCache transfer for disaggregated inference with dynamic scaling” — arxiv.org, Oct 31, 2025
“Perplexity open source garden for inference technology” — github.com, Nov 4, 2025
Source last verified Jul 26, 2026
Model Context Protocol
by Anthropic Agent-integration standard with an official server exposing Sonar and Search.
Receipts · 1
“The official MCP server implementation for the Perplexity API Platform” — github.com, Mar 10, 2025
Source last verified Jul 26, 2026
Receipts · 2
“Perplexity has been able to create their own search engine that delivers uncompromising quality” — blog.vespa.ai, Apr 15, 2025
“Perplexity has used Vespa.ai's managed platform to efficiently scale its RAG architecture” — vespa.ai, Apr 15, 2025
Source last verified Jul 26, 2026
Proprietary crawler and index feeding hybrid lexical and semantic retrieval.
Receipts · 1
“Our search index tracks over 200 billion unique URLs, with capacity to track many hundreds of billions more.” — research.perplexity.ai, Jul 17, 2026
Source last verified Jul 26, 2026
Grades search quality with LLM-as-judge evaluations over public benchmarks.
Receipts · 2
“We grade all benchmarks using the same prompted classifier methodology used in the original work” — research.perplexity.ai, Jul 17, 2026
“you will build specialized evals to improve answer quality across Perplexity” — jobs.ashbyhq.com, Jun 29, 2026
Source last verified Jul 26, 2026
Amazon SageMaker HyperPod
by AWS Early distributed training platform, since replaced by a self-managed GPU fleet.
Receipts · 2
“Amazon SageMaker HyperPod's built-in data and model parallel libraries helped us optimize training time on GPUs” — aws.amazon.com, Mar 2024
“Perplexity Accelerates Foundation Model Training by 40% with Amazon SageMaker HyperPod” — youtube.com, Mar 27, 2024
Source last verified Jul 26, 2026
NeMo
by NVIDIA Framework behind the post-training run that produced R1-1776.
Receipts · 1
“They used NVIDIA's NeMo 2.0 framework to fine-tune the model” — hyperight.com, Feb 24, 2025
Source last verified Jul 26, 2026
Curates multilingual post-training datasets with in-house experts.
Receipts · 1
“they built a dataset of 40,000 multilingual prompts to retrain the model” — hyperight.com, Feb 24, 2025
Source last verified Jul 26, 2026
Fine-tuned classifier detecting prompt-injection attacks in the Comet browser.
Receipts · 2
“fine-tuned a version of Qwen3-30B that can scan raw HTML and detect prompt injection attacks” — x.com, Dec 3, 2025
“BrowseSafe is designed to analyze complete webpages quickly without affecting browser performance.” — mpost.io, Dec 4, 2025
Source last verified Jul 26, 2026
External security firm stress-tests Comet's agentic browsing against prompt injection.
Receipts · 1
“Before launching their Comet browser, Perplexity hired us to test the security of their AI-powered browsing features.” — blog.trailofbits.com, Feb 20, 2026
Source last verified Jul 26, 2026
Receipts · 2
“They upload my screenshot to Cursor and then ask it to write a change to the Swift UI file” — aol.com, Jul 14, 2025
“made it compulsory” — entrepreneur.com, Jul 15, 2025
Source last verified Jul 26, 2026
Internal agents automate end-to-end analytics workflows for the data org.
Receipts · 1
“You'll join a data team that's already using AI across its work.” — jobs.ashbyhq.com, Jul 23, 2026
Source last verified Jul 26, 2026
