AIBriefsLive feed →

Tagged

Perplexity News

Releases, search features and business moves from Perplexity. Curated and summarized from dozens of sources by AIBriefs.

LaunchAI Models2 sources

Perplexity releases pplx-embed-v2-context embedding models

Open-weight vision-language embedding models in 9B and 0.6B sizes, both returning one int8 vector per document or per chunk. The 9B model sets a new state of the art on Context-bench, Q2D-Chunk and ConTEB, and leads dense retrievers on visual ViDoRe-V3.

AnalysisBusiness1 source

Perplexity CEO pitches AI as the "CFO of your life"

At Money 20/20 USA in Las Vegas, Aravind Srinivas described AI shifting from answering questions to running background tasks like tracking spending. His framing: "Everyone's the CEO of their own life. Perplexity is the CFO."

EventBusiness1 source

Cambridge Union hosts Perplexity CEO Aravind Srinivas

The Cambridge Union posted its October 3 conversation with Perplexity CEO and co-founder Aravind Srinivas, recorded in its Debating Chamber. The video runs as a full session with the AI search company's chief executive.

AnalysisAI Models3 sources

Perplexity and NVIDIA unveil JOLT post-training method

JOLT (Joint On-policy Learning and Teaching) matched GRPO's MATH accuracy with 6.5x fewer training tokens by training one model as both teacher and student. JOLT+ hit 72.09% on AppWorld state-test pass rate vs 62.20% for GRPO, though JOLT without student reward fell to 51.92%.

LaunchAI Agents6 sources

Perplexity open-sources SPACE, a Firecracker microVM sandbox for AI agents

SPACE gives each agent task its own Firecracker microVM, with credentials and outbound traffic controlled from outside the VM; code ships under Apache 2.0. The Rust codebase (release 0.8.0) cut median sandbox create latency from 185 ms to 60 ms versus Perplexity's previous provider.

LaunchAI Models7 sources

Perplexity open-sources Decider V1.2 decision models

pplx-decider-v1.2 is a 27B model at 2 cents per million input tokens; pplx-decider-flash-v1.2 is a 4B model built on Qwen3.5-4B at 1 cent, distilled from the 27B teachers. Output tokens are free, and both return typed answers with probabilities instead of text.

AnalysisPolicy1 source

Perplexity CEO: only AI you own is trustworthy

In a Sourcery interview clip, Perplexity's CEO argues the only trustworthy AI is one you own and that runs on your own device and network, including a version that can run fully air gapped.

AnalysisBusiness3 sources

Perplexity CEO: solopreneur spends ~$1M/year on the product

Aravind Srinivas says a one-person company spends upwards of $1M a year on Perplexity while making $3M a year by automating work. Perplexity is valued at ~$20B with more than $2B in reported funding, backed by NVIDIA, SoftBank, IVP and NEA.

LaunchAI Models4 sources

Perplexity open-sources pplx-embed-v2-late embedding models

Perplexity released pplx-embed-v2-late, ColBERT-style multimodal embedding models in 0.6B and 9B sizes on Hugging Face under the MIT license. The 9B model scores 92.4% on MADQA and beats the best dense model by 8.7pp on BrowseComp+; a 9B index can be queried with the 0.6B model at 63.5% on ViDoRe v3.

LaunchAI Models4 sources

Perplexity's pplx-decider-v1.1-27b tops Decision Bench at 94.5% accuracy

The open-weights multimodal decision model hit 94.5% accuracy across 1,071 cases at $0.017 per 1,000 decisions, per Perplexity. It returns typed probabilistic answers — yes/no, choice, or rubric score — for up to 128 questions per request, and is now on OpenRouter at $0.02/M input with free output.