Introducing Linkup Flash: the most accurate web search for agents, in under 250 milliseconds

Flash is available through one parameter:

const response = await client.search({
  query: "What changed in the latest EU AI Act guidance?",
  depth: "flash",
  outputType: "searchResults",
});

From search orchestration to intelligent retrieval

The first generation of web-augmented AI systems delegated most of the work to the search layer: query decomposition, retrieval planning, source selection, page reading, answer synthesis.

As AI systems take on mission-critical tasks, teams are bringing that work inside the application to keep control over the output. Planning, tool selection, iteration, and evaluation now happen in the agent itself: it decides what to retrieve and how each piece of evidence shapes its next action.

As orchestration moves into the application, the interface with search becomes narrower and more demanding: relevant, clean, AI-ready evidence, delivered in milliseconds.

The same shift is underway at the model layer. As reasoning and tool use improve, system quality depends less on what a model memorized during training and more on reliable access to the right information at inference time. Current knowledge cannot live in model weights alone.

Flash is built for this interface: reliable web evidence, as fast as we can serve it.

Built for high-throughput, low-latency applications

Latency bands are not points on a continuum. They determine where search can sit in a system's architecture. Around half a second, web search works as a tool: the agent decides to search, and the workflow pauses while it does. Below 300 milliseconds, search can move inside the loop itself — consulted at every conversational turn, every reasoning step, every training episode, without the system being designed around the wait.

Flash is built for this regime: the fastest search accurate enough to be trusted at every step.

• Real-time voice and embodied systems. Voice agents, wearables, vehicles, and robots interact in real time: complex workflows inside latency budgets of a few hundred milliseconds.
• Deep research. Deep-research systems work in rounds: search, read, decide what is still missing, search again. A single report can involve dozens of searches, and because each round depends on the last, retrieval latency compounds. Cutting it lets the system run more rounds in the same time — explore more leads, discard more dead ends, and read the best sources more carefully.
• Reinforcement learning. Training agents with reinforcement learning means running thousands of practice episodes in parallel, each making its own searches. When every episode waits on retrieval, search latency sets the pace of training itself. Faster calls and higher throughput means greater velocity, which compounds.
• Small language models. Small models trade stored knowledge for reasoning and tool use. That trade only works if external knowledge arrives fast enough to preserve what makes them attractive: speed and cost. Flash supplies current information at inference time, at a latency that keeps small and local models viable for knowledge-intensive tasks.

Low latency starts at the index

Flash holds retrieval quality at this latency because it runs on Linkup's own index and search infrastructure: we control the full path from source selection and refresh to document processing, retrieval, and ranking. The speed comes not from any single component but from optimizing storage layout, retrieval stages, and ranking models together, and evaluating them under the conditions they face in production.

This efficiency makes the stack portable — across regions, inside a customer's cloud, or on-premises — so we can serve the same ultra-low latency to clients anywhere in the world.

How we measure performance

We benchmarked Flash against the fastest search mode of each major web search API on SimpleQA, under identical conditions: the same queries, the same number of results, latency measured client-side, each provider called from the cloud region closest to its own infrastructure. Latencies measured this way can exceed providers' published figures, which are typically produced under conditions each provider chose.

Flash returns results in 250 milliseconds at the median and 311 milliseconds at the 90th percentile — the most accurate web search API on SimpleQA under 300 milliseconds.

Search systems designed around the application

Most web search APIs expose a fixed system: one index, one ranking pipeline, one deployment model. We think the knowledge layer of an AI system should be designed around the application it serves.

Operating our own index makes that practical:

• Client-specific indexes and retrieval pipelines. The source corpus, refresh policy, ranking models, retrieval stages, and query capacity can each be configured per application.
• Deployment that follows security and privacy requirements. Zero Data Retention, regional processing, private cloud, and on-premises installations.

Teams control both what their AI can retrieve and where that retrieval takes place.

Try Linkup Flash

Flash is available today in the Search API:

depth: "flash"

Try Flash in the playground, read the Search API documentation, or talk to our team about dedicated capacity, a tailored index, or a private deployment.