May 2026 Summaries

Perplexity has expanded its Computer assistant into Microsoft Word, Excel, PowerPoint, and Outlook following its earlier Microsoft Teams integration, allowing paid Pro, Max, and Enterprise customers to perform AI-assisted work directly within Microsoft 365 applications. Available through add-ins and side panels, Computer can draft and edit documents, analyze spreadsheets, create presentations, compose emails, prepare meeting materials, and draw on web research, connected files, SharePoint, Teams, Outlook, Salesforce, FactSet, Snowflake, and more than 400 additional apps and sources. Perplexity says the tool automatically selects among more than 20 models for tasks and provides citations for answers and figures to support verification. The company highlights enterprise controls including SAML single sign-on, audit logs, connector and model management, and a policy that enterprise data is not used to train models, while users retain the ability to review actions before they are completed.

Perplexity has open-sourced Bumblebee, a read-only Go-based scanner for macOS and Linux developer endpoints that helps security teams identify exposure to risky software packages, editor and browser extensions, and AI agent configurations during supply-chain incidents. The tool uses organization-maintained catalogs of vulnerable components, which can be updated from public disclosures, intelligence feeds, or internal research through a reviewed workflow, then scans endpoints using baseline, project-specific, or incident-response deep profiles. Unlike repository-focused SBOM and vulnerability scanners or application inventory tools, Bumblebee examines local developer environments across language package managers, MCP configurations, VS Code-family extensions, and Chromium- and Firefox-based browser extensions. It reads metadata such as manifests, lockfiles, and installed package information without invoking package managers, executing code, reading source files, or monitoring network and process activity, reducing the risk that scanning itself triggers malicious install-time scripts.

A performance-focused reimplementation of the XLM-RoBERTa Unigram tokenizer substantially reduced CPU-side inference overhead by replacing allocation-heavy and pointer-chasing components in Hugging Face’s Rust tokenizer with a specialized zero-allocation Viterbi encoder. Unigram tokenization selects the highest-scoring segmentation through dynamic programming, but profiling showed the reference implementation spent significant time creating strings, performing hash-map lookups, allocating memory for each token match, and traversing heap-scattered trie nodes. The new implementation first eliminated redundant allocations and stored token IDs and scores directly in trie nodes, then replaced the hash-map trie with a cache-friendly double-array trie, added bitmap-based child validation and cache-line-aligned node packing, and used 2 MB huge pages to reduce translation lookaside buffer overhead. At production-scale inputs, the resulting encoder achieved roughly 5× lower median latency than Hugging Face tokenizers, about 2× faster performance than native SentencePiece, and about 1.5× faster than IREE, while reducing instructions per encode from 3.66 million to 1.04 million and preserving token-exact output. In deployment, these improvements cut preprocessing CPU utilization by 5–6× and removed double-digit milliseconds from reranker latency, illustrating that tokenizer design, memory allocation, cache locality, and wrapper overhead can materially affect performance for small-model inference workloads.

Perplexity describes a query-aware extractive context-compression system designed to remove irrelevant material from retrieved search results before they reach answer models, improving accuracy, latency, and cost while preserving source wording for reliable citations. Rather than generating summaries, the system selects the most useful spans from each query-document pair, using a bidirectional pplx-diffusion encoder and a lightweight compression head trained on 750,000 pairs labeled through a two-stage LLM judging pipeline. A distilled 17-layer production model matched the quality of a 28-layer teacher while reducing inference latency by 35–40% and GPU compute by 40–45%, with p99 serving latency below 20 milliseconds. Tests on BrowseComp and SimpleQA found that compression reduced token usage by 10–70% while improving accuracy, and an analysis of 1,000 queries showed a 63% increase in vital content alongside declines in irrelevant navigation, metadata, advertising, and duplicate text. The approach gives systems and API users a way to adjust compression according to desired tradeoffs among evidence recall, speed, and cost.

Perplexity has introduced a Snowflake App Connector that allows Pro, Max, and Enterprise users to query enterprise data through its Computer product using plain-language questions, aiming to reduce reliance on SQL expertise, dashboards, and data-team request queues. The connector can support tasks such as identifying support-ticket drivers, forecasting demand, segmenting customers, and evaluating marketing performance, while a shared semantic layer maps business definitions, metrics, tables, and common questions to appropriate SQL. The product draws on Perplexity’s internal Snowflake Slackbot, which employees reportedly use for up to 3,000 queries per week across functions including product, analytics, growth, support, and security. Security features include encryption, privacy protections, Snowflake role-based access controls, optional read-only permissions, and User OAuth so individuals can access only data already permitted through their Snowflake accounts; Perplexity states that connected data is not used to train models. Administrators can generate and review a data map through the Connectors settings, a process that may take up to 90 minutes depending on the platform’s complexity.

Perplexity Computer is an autonomous agent that can write and execute code, browse the web, and access external services, built on Perplexity’s SOC 2 Type II-attested infrastructure and enterprise security features. To manage the added risks of executing code and interacting with live services, it uses Firecracker microVMs with dedicated kernels, isolated filesystems and networks, short-lived credentials, encrypted communications, and separation between storage and execution environments. Connector access is controlled by administrators and authenticated by users, with minimal-data transmission, encryption for file data, and a policy that enterprise task data and attachments are not used for model training, while attachments are deleted after seven days. Its prompt-injection defenses combine the BrowseSafe detection model, external-content classifiers, tool-specific guardrails, untrusted-content labeling, and safe stops for suspicious material, with ongoing updates informed by audits, red-team testing, and bug reports. Enterprise administrators can further govern access, connectors, models, billing, and monitoring through audit logs that integrate with SIEM platforms such as Splunk, Azure Sentinel, and Datadog.

Perplexity describes adapting its inference engine to serve post-trained Qwen3 235B mixture-of-experts models on NVIDIA GB200 NVL72 racks, whose 72 Blackwell GPUs combine large HBM capacity, 72-way NVLink connectivity, NVLink Switch SHARP reductions, and InfiniBand links to support high-throughput, lower-cost model deployment. The system separates computationally intensive prompt prefilling from memory-bound token decoding, using tensor and expert parallelism across four GPUs for prefill and data plus expert parallelism, scaled up to 16 GPUs, for decoding; KV caches are transferred flexibly between these differently sharded stages. Custom CuTeDSL kernels and NVLink-based communication improve all-reduce and MoE dispatch-and-combine operations, while SHARP reduces communication overhead and reportedly lowers prefill and decode latency by about 10 percent in practice. Compared with Hopper-based deployments, Blackwell’s greater compute capability, memory bandwidth, and wider NVLink domain enable better scaling beyond eight GPUs, with decoder sharding across 16 ranks justified by faster MoE matrix operations despite modest added communication latency. The deployment uses MXFP8 weight quantization and speculative decoding to improve performance while preserving acceptable accuracy, while more aggressive 4-bit formats remain under consideration because of their accuracy impact without quantization-aware training.

Perplexity has introduced Personal Computer in its new macOS app, extending its Computer agent system from cloud-based workflows to local Macs and Mac minis. The tool can autonomously work across local files, native Mac applications, web tools, more than 400 connectors, and Perplexity’s secure server-side development sandbox, while users retain control through approval requests for important actions. Designed for tasks that combine documents, spreadsheets, browser content, notes, and other materials, it can organize project files, compare documents, create drafts and reports, analyze data, and build live dashboards. A Mac mini is presented as an always-on option for continuous agent activity that can be initiated or approved from other Apple devices, while the Comet browser enables web-based operations without direct connectors. The new app is available for download outside the Mac App Store to all users, with Pro and Max subscribers using subscription-linked credits, and it will eventually replace the previous Perplexity Mac app.

Perplexity has launched Finance Search for its Agent API, a tool that combines licensed financial datasets, real-time market information, and cited web sources in a single call to support financial research workflows. Designed for time-sensitive tasks such as valuation lookups, earnings recaps, disclosure monitoring, and tracking analyst estimate changes, it retrieves prices, company fundamentals, transcripts, and market context through a consistent schema without requiring developers to integrate individual data providers. The company says Finance Search routes requests to relevant structured financial sources, enabling agents to generate answers, charts, memos, or longer research workflows using current data. On the FinSearchComp T1 benchmark, Perplexity reports that the tool delivered the highest and most consistent accuracy for live financial-data queries, while also achieving the lowest cost per correct answer by reducing reliance on large volumes of generic web text. Results include inline citations for source verification, and developers can select models, monitor token usage, and use benchmark-tested configurations.

Perplexity’s in-house Runtime-Optimized Serving Engine, or ROSE, serves embedding models, LLMs, ranking systems, and other transformer workloads across NVIDIA Hopper and Blackwell GPUs, emphasizing rapid adaptation to new model architectures alongside high inference performance. Although ROSE initially relied on Triton and uses CUTLASS and cuBLAS for matrix multiplication and attention, it has adopted NVIDIA’s Python-based CuTeDSL as its main environment for many specialized GPU kernels because its just-in-time compilation, low-level hardware access, MLIR-to-PTX pipeline, and layout abstractions enable aggressive specialization without CUDA’s lengthy template-compilation burden. The system differentiates prefill workloads, which prioritize compute throughput over many tokens, from decode workloads, which are typically memory-bandwidth and latency constrained, adapting warp counts, grid assignments, vector widths, and synchronization strategies accordingly. CuTeDSL has supported optimized QK normalization kernels that reportedly improve throughput and latency by two to three times, as well as mixture-of-experts dispatch and combine kernels for local, tensor-parallel, NVLink data-parallel, and InfiniBand deployments, including a reported twofold improvement over previous CUDA kernels for NVLink token transfer and a further latency gain for peer-to-peer kernels. The DSL also eases Blackwell support through configurable SM counts, wider memory instructions, and newer quantization formats, while internally developed CuTeDSL code patterns have gradually made AI-assisted kernel generation more useful despite models’ continuing difficulty with low-level synchronization and vectorization. Perplexity concludes that CuTeDSL approaches CUTLASS C++ performance while improving development speed and configurability, positioning ROSE to further fuse computation, communication, normalization, and activation operations across its inference workloads.

Perplexity has launched Premium Health Sources to provide users with cited, clinically rigorous medical information from professional resources often restricted behind subscriptions, responding to the significant share of health-related queries on its platform. Automatically activated for relevant questions in Perplexity Computer, the feature incorporates sources including the New England Journal of Medicine, BMJ Journals, and BMJ Best Practice, offering peer-reviewed research, clinical studies, and point-of-care guidance for consumers as well as healthcare and biopharma teams. Planned additions include Micromedex, EBSCOhost, Health Affairs, VisualDx, specialty guidance from medical associations, and Springer Publishing materials. The initiative is part of Perplexity’s broader health strategy, including Perplexity Health and its Health Advisory Board, aimed at making reliable health information more transparent and accessible.

Perplexity has launched Computer for Professional Finance, a version of its platform designed to support finance teams’ research, analysis, and decision-making workflows with cited and traceable data. Users can connect licensed sources such as Morningstar, PitchBook, Daloopa, and Carbon Arc through MCP connectors or use built-in tools drawing from 14 providers, while the platform’s memory retains source preferences for future tasks. It includes 35 workflows across ten finance segments, enabling teams to create outputs such as company tearsheets, private-equity sourcing dashboards, and annotated stock charts that combine financial data, events, and news. Each numeric result links to its underlying source, including SEC filing pages and calculation details, so supporting evidence remains attached to charts, memos, dashboards, and slides. Computer is available through Microsoft Teams and the Agent API, with an Excel add-in planned, and its Finance Search Tool provides live equity, cryptocurrency, and foreign-exchange data.

Perplexity has launched its Computer AI agent orchestrator for Microsoft Teams, extending a product it says has completed more than $2.8 billion in labor-equivalent work for paid subscribers within a month. The integration lets users assign tasks through direct messages or by tagging Computer in shared Teams conversations, where it can use existing context, access workspace knowledge, read threads, send messages, and keep requests and results visible to collaborators. Computer connects with Microsoft services including Outlook, OneDrive, SharePoint, Excel, Azure DevOps, Dynamics 365, and Word, while an Excel side-panel integration remains in beta. It can research the web, analyze data, create documents, reports, presentations, applications, and websites, and run recurring background workflows such as market scans or support-ticket processing. Perplexity also introduced data App Connectors that allow approved users to query sources such as Snowflake and Databricks in natural language, a capability aimed partly at executives, 92% of whom the company says use Computer weekly. Computer for Teams is available through the Microsoft Marketplace.

Perplexity’s guide to Agent Skills describes them as modular, progressively loaded packages of context, instructions, code, references, and assets that help agents perform specialized tasks more reliably than general model knowledge alone. Unlike conventional software, effective Skills prioritize dense, high-value guidance, carefully curated hierarchy, conditional information loading, and detailed “gotchas” that address known failure modes, while avoiding obvious instructions the model already knows. Each Skill’s concise description is especially important because it determines when the agent loads it, and poorly scoped routing can cause unwanted overlap with other Skills. The guide recommends determining whether a Skill is necessary through real-task testing, writing evaluations before implementation, using positive and negative routing examples, placing heavy or conditional material in auxiliary files, and iterating extensively before release. Ongoing maintenance should focus on evaluated routing changes, accumulating failure-derived guidance, progressive-loading checks, end-to-end task evaluations, and testing across model families. Although Skills can automate recurring workflows and encode domain expertise or organizational preferences, the guide emphasizes that every Skill consumes context and requires careful curation, making brevity, specificity, and necessity central design principles.