For years, Tailscale has worked on making internet connections faster. Now the company says it has a plan for moving that work ahead, and it centers on keeping packets small while giving hardware more than one path to travel.
Today’s announcement covers a range of upgrades meant to increase output across app connectors, subnet routers, and exit nodes. Among the additions are multi-queue technology, reduced packet sizes, and quicker startup times, with certain gains arriving later in 2026.
Packet Shipping and the Big Buffer Problem
The packets that pass through most networks tend to be quite small, roughly 1 KiB in size. Yet the wireguard-go code used by Tailscale’s current setup forces it to treat them as if they were much larger. The company’, which manages encryption and networking, provides a single buffer size for unpacking those packets: 64 KiB.
This implies that a 1 KiB packet is copied into its very own 64 KiB buffer each and every time. Instead of copying these packets elsewhere, Tailscale now lets them stay put on Linux and Android. The system identifies where each packet begins and ends within the single large read rather than making a fresh copy of it.
Small packets tend to remain small in memory, and numerous of them can share a single allocation. Because of this, they spend less time being copied. According to the company, this arrangement produced a roughly 5% speed-up across many network configurations.
Shorter Queues and Their Savings
Between stages of the pipeline, packets wait in queues designed to absorb bursts of traffic. Testing revealed that much of that capacity sat idle. Trimming queue depth cut waiting time and reduced memory overhead alike.
The extra memory that became available was put to use by helping two of the busiest node types: subnet routers and app connectors, which received the benefits from Tailscale’s savings.
Multi-Queue Lanes for Subnet Routers
A small home lab network may manage just a handful of non-Tailscale devices, while a subnet router sitting before a cloud deployment serving hundreds of peers handles far heavier traffic. The two setups can look entirely distinct from one another.
For a long time, subnet routers, app connectors, and exit nodes handled several separate streams through one ordered, single-thread pipeline. Each connection had to share that single lane, since a receiving application must never see its own packets arrive out of order.
Tailscale gained room for a multi-queue system thanks to its smaller memory use, with multiple lanes instead of a single one, scaled to fit the machine’s resources rather than tied to the number of peers. Packets stay on their assigned lanes, which operate at the same time, so work can be divided among CPU cores.
“This translates into lower latency, essentially faster processing of data from the moment we read it off the wire to the moment we send it to the OS,” said Alex Valiushko, member of technical staff at Tailscale.
Higher Throughput With writev
Tailscale uses Linux’s writev feature to hand several chunks of packet information to the kernel all at once, instead of copying and joining those pieces separately first. The v in writev refers to “vector”, which means Tailscale can point to distinct pieces of data that need to be handled, without actually moving them.
The result is a smaller number of stored packet copies, less frequent writing to memory, and a faster rate of information passing through the system.
Faster Startup With netmap Caching
Only Linux and Android benefit from these current speed-ups. Tailscale has been developing features for other systems, however, and clients will soon be able to take advantage of netmap caching to start more quickly under many conditions.
When a device links up with Tailscale, it normally begins by reaching out to Tailscale’s control plane, which takes about 100 milliseconds on a standard network. Afterward, the device proves its identity and receives a “network map”, or netmap, which lays out the other devices it can connect to and the routes for getting there. Under normal conditions, the whole startup sequence moves quickly, close to instant.
The connection can drag when the airplane’s Wi-Fi is weak, or when a hotel applies heavy filtering to its network.
What These Changes Support
This has rendered Tailscale suitable for workloads that demand greater performance, which makes it possible to apply Tailscale to continuous integration, agentic workflows, remote development environments, robotic edge devices, heavy data and telemetry workloads, and beyond.
| Workload | Example |
|---|---|
| Continuous Integration | Automated testing pipelines |
| Agentic Workflows | Remote orchestration tasks |
| Remote Development Environments | Cloud-hosted coding sessions |
| Robotic Edge Devices | Factory automation systems |
| Heavy Data and Telemetry | Large-scale monitoring |
The Road Ahead
Tailscale continues to address performance tooling problems it hopes to resolve for its users. The firm has dedicated resources toward improving Tailscale’s speed, beginning with enhanced TCP throughput on Linux machines, followed by exceeding 10Gb/s on bare metal using wireguard-go, and subsequently employing segmentation offloads to push UDP-based application throughput past 4x.
The company still aims to reach its original target: finding a way through NAT Traversal, even when networks refuse to cooperate. The data plane must hold steady no matter what.
Key Facts
- 5% speed-up from smaller packet handling on Linux and Android
- Multi-queue system lands in the second half of 2026
- Netmap caching coming soon for faster startup
- Wireguard-go surpassed 10Gb/s on bare metal
- Throughput improved over 4x for UDP-based apps with segmentation offloads
The firm addresses the issue forthrightly and specifies its remedy plainly. Its assertions rest upon genuine statistics and concrete illustrations.
Whether the updates arrive as advertised depends on whether the promised 5% speed-up and reduced latency stand up to real-world deployments.
The path forward features a series of stable client releases carrying enhancements to throughput and memory overhead, alongside expanded performance tooling for customers.
Source material: “Making Tailscale Faster,” tailscale.com.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.