Is Perplexity AI Down?

HYBRID COMPUTE ON MAC -- CLOUD AND LOCAL MODELS SPLIT A SINGLE TASK, WITH AN ON-DEVICE PRIVACY GATE (2026-09-01, vendor-primary; this page was last reviewed 2026-08-20 and had only a vague June note about a 'hybrid local-server inference orchestrator', which this supersedes): Perplexity shipped the feature that makes Personal Computer defensible for confidential work. **WHAT IT ACTUALLY DOES:** in the newest Perplexity for Mac app, users 'can now add a compact local model post-trained for Perplexity Computer and choose to run tasks on device'. Each task is **divided between the cloud and the local model** rather than routed wholly to one: 'The cloud handles frontier reasoning, web search, and planning, while the local model on the Mac processes private files, sensitive information, and on-device actions needed at every step.' Computer coordinates the whole workflow, so the user does not have to split prompts manually. **THE PRIVACY GATE IS THE PART WORTH READING CLOSELY, AND IT IS MORE THAN AN ON/OFF SWITCH:** 'A privacy gate on the Mac controls what may leave the machine. Before information from a protected file reaches the cloud, the privacy gate can **mask sensitive details, keep it local, refuse the action, or ask for consent**.' Four graded outcomes, enforced on device -- that is a materially stronger claim than 'we do not train on your data', because it is a mechanism rather than a promise. **REMOTE USE:** 'Perplexity Computer works with iPhone' -- prompts can be sent from anywhere while the local model keeps running on the Mac. **HARDWARE FRAMING:** Perplexity names Apple's new Mac mini and Mac Studio, 'announced last week', as ideal platforms, leaning on Apple silicon and unified memory. **THE HONEST LIMITS:** this is **Mac-only** -- the Windows version announced around 2026-06-04 still has not shipped -- and the local half is a **compact** model, not a frontier one, so the cloud is still doing the hard reasoning. The privacy guarantee covers what the gate is configured to protect, not everything you touch

LILY: PERPLEXITY BUILT ITS OWN APPLE-SILICON INFERENCE ENGINE, AND PUBLISHED THE NUMBERS (2026-09-01, vendor engineering post): the local half of Hybrid Compute does not run on Apple's standard stack. Perplexity wrote **Lily, a lightweight local inference engine built specifically for Apple silicon and Qwen3.6-35B-A3B**, with separate optimizations for prefill and decode. **ARCHITECTURE:** a **Rust runtime** loads the checkpoint and manages session state and the generation loop, an **OpenAI-compatible chat-completions API** accepts requests and streams tokens, and **custom Metal kernels** execute Qwen-specific operations -- **'Neither PyTorch nor MLX is in the execution path'**. The rationale is specialization: MLX-LM's reusable kernels must support many architectures, whereas Qwen3.6-35B-A3B is a sparse hybrid (MoE routing plus fixed-size recurrent states and full attention) whose irregular workloads reward an engine that coordinates kernels, data movement and scheduling around one fixed structure. **VENDOR-PUBLISHED PERFORMANCE, ON A MAC STUDIO WITH A 40-CORE GPU AND 128 GB OF UNIFIED MEMORY:** across ten prompt lengths for prefill and ten context lengths for decode, from 256 to 128K tokens, Lily averages **1.23x MLX-LM's prefill throughput and 1.35x its decode throughput**; at a 4K-token prompt and 4K-token decode context it reaches **5,749.9 prefill tokens/sec and 186.6 decode tokens/sec**. **TREAT THESE AS FIRST-PARTY:** they are Perplexity's own measurements against MLX-LM on one machine configuration, with no third-party replication -- and a 1.23-1.35x gain over a general-purpose framework is a solid engineering result, not a step change. **A standalone demo is publicly available on GitHub**, which is unusual enough to be worth noting: the engine underpinning a paid product is inspectable

COMPUTER IN EMAIL -- THE AGENT NOW RUNS FROM YOUR INBOX, AND IT IS LIVE FOR EVERY COMPUTER USER TODAY (2026-08-18, vendor-primary): Perplexity added email as a launch surface for Computer. You '**send, forward, or cc computer@perplexity.com**' -- a new message, a forwarded thread, or a cc on an existing conversation -- and '**Computer verifies the sender, reads the full thread and any attachments, and runs the task with the sender's own connectors and permissions**'. The reply comes back **in the same thread**, with deliverables like **Excel models, PDFs and decks returned as attachments**. **AVAILABILITY IS THE STRONG PART:** '**Computer in Email is available to all Computer users today**' -- no waitlist, no tier gate beyond already having Computer (Pro $20/mo and Max $200/mo per the pricing above). **THE SECURITY MODEL, WHICH IS THE FIRST THING TO ASK ABOUT AN EMAIL-TRIGGERED AGENT:** Perplexity says Computer **verifies the sender before answering**, so 'a task only runs with that person's own access', and for security it '**replies only to the sender**' -- **'Reply all' support for Enterprise is 'coming soon'**, i.e. not shipped. Every email task is a full Computer session viewable on web and mobile '**with the same steps and audit trail**' as a web-app task. **CONTEXT CARRIES OVER:** an email task draws on the same **Brain** self-improving memory as web and Slack sessions, so team terminology, output-format preferences and prior decisions apply. **WHERE IT SITS:** email joins **Computer in Slack** and **Computer in Teams** as the third collaboration surface. Perplexity's rationale is that 'nearly a third of knowledge work still happens over email' and its worked examples are pointedly finance and legal -- forwarding a CIM and data-room thread to get a first-pass LBO model back as Excel, cc'ing Computer on a redline exchange for a summary of open issues, emailing for a five-year DCF. **THE CAUTION WORTH STATING:** a forwarded thread hands the agent everything in that thread including attachments, and the trigger is an email address -- so sender verification is the entire boundary between 'my agent' and 'anyone who can spoof me'. Perplexity does not describe the verification mechanism in the post

FIVE-PART RELEASE: PROJECTS, MODEL COUNCIL, US-HOSTED KIMI K3, VOICE MEMORY AND AN OPEN-SOURCE AGENT SECURITY TOOL (2026-08-04, vendor changelog): Perplexity's largest single release since Computer launched, and none of it was picked up by aggregator coverage. **(1) PROJECTS REPLACE SPACES.** Projects are shared Computer workspaces with a **persistent file system where humans and agents edit the same files**, a self-improving memory via **Brain** that carries context between tasks, and **connectors scoped per project** rather than per account. **Every existing Space is migrated automatically** -- if you had Spaces, they are now Projects, so this is a change you get whether or not you wanted it. Collaboration detail that matters for privacy: invited teammates share files, artifacts and history, but '**personal memory and connector accounts stay scoped to each individual**'. You can link a Slack or Teams channel to a project with `/project` and start tasks in-channel with full project context. **(2) MODEL COUNCIL IN COMPUTER** -- assemble '**a board of top models for a single question**': pick **two to eight models across OpenAI, Anthropic, Gemini and open source**, choose an analysis depth, and Computer runs each independently then synthesises **where they agree, where they disagree, and what each uniquely surfaces**, before turning the result into reports or decks. This is the most interesting product idea in the release: it sells model disagreement as the deliverable, which is the opposite of every router that tries to hide it. **(3) KIMI K3, US-HOSTED** -- the 2.8T open-weight frontier model is now in Perplexity and Perplexity Computer for **Pro and Max**, '**hosted exclusively on US-based servers**', with 1M context and native vision, explicitly pitched as getting the model '**without data leaving US infrastructure**'. That is a jurisdictional answer to the main enterprise objection to Chinese models, made possible by the weights being public (see /tools/kimi). **(4) VOICE MODE MEMORY** -- voice on web and Computer now opens with your personalisation context (profession, interests) already loaded, so you stop re-explaining yourself each session. **(5) NUMBAT** -- a genuinely unexpected one: Perplexity released an **open-source agent-detection and response layer** giving security teams live visibility into what AI coding agents do on endpoints, **with the ability to block risky actions before they run**. It ships as a **single Go binary for macOS, Linux and Windows**, integrates with the leading agent harnesses, emits **hook-based and OpenTelemetry telemetry**, enforces **CEL rules** to stop network egress, sensitive file access or shell commands on-device, and can **reconstruct a forensic timeline from on-disk artifacts even with no live telemetry running**. A search company shipping open-source EDR for coding agents is a real signal about where the agent-security gap is

PERSONAL COMPUTER FOR WINDOWS (2026-07-28, vendor-primary): Perplexity's agent platform came to Windows -- 'Today we're launching Personal Computer for Windows, bringing Perplexity's agent platform to more than a billion devices globally.' It works on local files (create and edit Word/Excel, find files and file them in Explorer), moves between local files and the web, connects to Microsoft 365 (Word, Excel, PowerPoint, Outlook, Teams) plus 400+ App Connectors, supports voice mode, and hands tasks off from phone to desktop. **AVAILABILITY CORRECTION -- this is NOT Max-gated, contrary to some coverage.** The vendor page states plainly: '**Personal Computer is available now for Pro, Max, and Enterprise subscribers on Windows 10 and 11**' (verified by direct fetch 2026-08-03). Safety posture: it 'always alerts a user before it takes a sensitive action like sending an email or deleting a file', files are created in a secure sandbox, and actions are auditable. Perplexity's own scale claim: Computer has performed 'more than $9.4B in labor-equivalent work' since launching earlier this year. This closes the Mac-only gap the April 2026 launch left open

BRAIN MEMORY + CLAUDE OPUS 5 UNDERNEATH (2026-07-13 and 2026-07-27, vendor changelog): **Brain**, Perplexity's self-improving memory system, launched in Computer on 7/13 -- it builds a private context graph across sessions, connectors, files and past decisions, refreshes nightly, links every memory back to its source, and can be switched off. Vendor-reported effect: it 'lifted answer correctness 25% and recall 16% while cutting cost 13% on tasks with prior context.' On **7/27 Brain went global for Max subscribers** in 15 languages. Same day, **Claude Opus 5 went live across Perplexity Search and Computer** -- vendor wording: 'On the WANDR benchmark, it outperformed every model tested except Fable 5, while costing 57% less per task.' Also in this window: mid-task orchestrator model switching, one-click website publishing to pplx.app or your own domain via Vercel, and Agent API Skills. Treat the Brain and WANDR figures as first-party