a few open source contributions from perplexity recently: •pplx-decider-v1-27b: SoTA multimodal decision model. 85.7% average across 11 benchmarks, ahead of Jev. •pplx-embed-v2-context-9b-preview: SoTA contextual embeddings, best on ConTEB and turbopuffer context-bench. •Lily: local inference engine for Apple silicon. rust plus custom metal kernels, no pytorch or mlx. 1.23x faster prefill and 1.35x faster decode than MLX-LM on an M5 Max •PII-Tracer: 0.6B on-device PII classifier that decides when a hybrid compute task stays on your Mac. beats OpenAI’s Privacy Filter on all 5 public benchmarks. also released with the PII-TRACE benchmark: 13k conversations in 13 languages •WANDR: benchmark for wide and deep research agents. 500 tasks needing 170k source-backed records •Numbat: agent detection and response for laptops and workstations. 52 rules, single go binary for macOS, linux and windows •Bumblebee: read-only supply chain scanner for dev machines. covers packages, MCP configs, and editor and browser extensions. a lot more open source contributions coming soon!

Try “Visualize….” in Computer and ask for whatever you want to learn visually. You will now get an inline viz with rich educative and interactive widgets and animations! Have fun! Best experienced on Standard or High effort.

Here are 5 examples of inline visualizations that Computer can create directly in your thread: 1. Show me how a jet engine works in an inline 3D cutaway.

Our goal is to vertically integrate our agentic infrastructure: own our sandboxes, and optimize the sandbox for the best silicon. Vera is far better than x86. Look forward to sharing more soon as we begin to rollout deployment of Perplexity Computer on Vera.

Your creativity is the limit in terms of what you can have Computer do for you. This is the kind of project that you can let an agent run for hours to come back with something great.

We're looking forward to working with the new @NVIDIA Vera CPU on SPACE.

We would love to hear how we can improve our developer platform!

Alert (again): Perplexity also just dropped a Jev alternative on Hugging Face 🚨 huggingface.co/perplexity-ai…

We’re open sourcing a state of the art multimodal Decision Model, pplx-decider-27b, and are offering it in a new Decisions API at 4 cents per million input tokens and free output tokens. We intend to bring down the price even further over the coming days. Enjoy!

Introducing the Perplexity Decisions API. It's powered by pplx-decider-v1-27b, our multimodal decision model trained to output a probability distribution over a fixed set of answers instead of text. It costs $0.04/million input tokens and scores 85.71% across benchmarks.

We’re rolling out inline charts, diagrams, visualizations in Computer. Including contextual financial charts from @tradingview

Seedance inside Computer is quite incredible for marketing and brand teams!

Today we’re launching a library of Amex-curated, ready-to-use skills in Perplexity Computer to eligible U.S. @AmexBusiness Small Business Card Members. The skills are pre-built workflows that handle everyday business tasks, like forecasting cash flow or generating marketing campaign. Small business owners already do the work of whole teams. They can put AI to work by choosing a task in Computer and adding details about their business, without having to write instructions from scratch. Learn more: pplx.ai/amex

I’ve been convinced contextual embedding models were dope since voyage-context-3 was released. Started using them and benchmarking them, realized that, esp for long documents, I needed them to rank not just the answer chunk, but the disambiguating context chunks highly. Done well, this could allow you or your search agent to read only the relevant fraction of the document rather than every page of a long document. So I started trying to measure that too. Offered to share my benchmark results with the perplexity team sometime after they released their contextual model. It’s worth mentioning - not all modeling teams want feedback from third parties like me. The perplexity team was eager for it and within a week had a new model for me to benchmark. So I’d run their new checkpoint on my benchmark, stare at the results, get a little frustrated that the benchmark was imperfect and wasn’t measuring everything as well as I wanted it to, iterate on it a bit and give them new results. We ran that back many times over the last few weeks. I’ve re-worked some part of the benchmark at least 9 times now looking over the results. Instead of getting frustrated with my moving target of a benchmark they just kept improving their model and refining their approach until they were clearly on top. Notably, they didn’t achieve this via privileged access to the benchmark, they simply iterated until they had a good model. Was really a pleasure working with @ESL_Sarah, @bo_wangbo, Markus, @antoine_chaffin, Louis, and Max. For those who aren’t going to read the blog I’ll share more about how “evidence recall” works soon.

we maintain several internal benchmarks to guide our customers toward better search relevance @perplexity's new model tops context-bench, our internal contextual embedding benchmark, and boosts document recall@10 by 50%+ over traditional SOTA embedding models

We’re open sourcing our state-of-the-art contextual embedding models, which perform best in turbopuffer’s context-bench.

We're opening access to anyone (no Perplexity account needed), and we will run all tasks delegated via email for free for a limited time. Forward or CC computer@perplexity.com to get started, and the agent will complete the task in the background while keeping the email context.

The only step that remains is project this while wearing a pair of glasses.

Computer can create interactive 3D diagrams to help you visualize things like furniture assembly.

The elite L/S analysts and growth investors I know from my gs days aren’t doing Q&A with AI. It’s all scheduled or event-triggered tasks. Market close? 15-20 automations activated. New filing or X post by CEO? 3-5 automations triggered. The amount of domain space they can cover and analyze with <5 minute latency is mind boggling.