> ## Content Index
> Fetch the complete content index at: https://www.tao.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# URnetwork’s First Apex Round Cuts p95 Request Time 63% in Testing
- URL: https://www.tao.media/urnetworks-first-apex-round-cuts-p95-request-time-63-in-testing/
- Published: 2026-11-10T16:26:41.000Z
- Updated: 2026-11-10T16:26:41.000Z
- Description: The winning code becomes the next round’s baseline after outperforming the network’s latency target.
- Author: Tristan Hillerich
- Tags: UR, Apex, Subnets, News

URnetwork’s first latency competition on Apex, Bittensor’s Subnet 1, has produced winning code that reduced p95 request time from 19.6 seconds to 7.1 seconds in the network’s tests. URnetwork reported a 63% reduction and said larger downloads also ran about twice as fast in its [Round 1 announcement](https://x.com/urnetwork/status/2119878753727701246?ref=tao.media).

The winning code is now the starting point for Round 2\. The competition uses successive rounds to improve the UR protocol, with each winner setting the baseline that the next field of entrants must beat.

[Apex said](https://x.com/Apex%5FSN1/status/2119893388593934578?ref=tao.media) URnetwork had sought a 20% reduction in p95 latency and that 14 independent solutions exceeded that goal. It put URnetwork’s monthly user count at 350,000\. The competition targets a concrete networking problem, the long waits that make internet connections feel unresponsive.

## Measuring the Slow End of a Connection

The p95 metric is the request duration at the 95th percentile: roughly 95% of measured requests finish at or below it, while the slowest 5% take longer. Unlike an average, it focuses attention on the slower end of the experience, where delays can remain noticeable even when most requests complete quickly.

Moving that threshold from nearly 20 seconds to just over seven seconds substantially shortens those waits in the test workload.

UR is an open-source protocol for network infrastructure. Its [research program](https://ur.xyz/research?ref=tao.media) covers the mechanisms that select network providers, distribute traffic across connections, and deliver data reliably. The system supports multiple IP addresses and transports, and researchers can optimize several interacting parts of the connection.

![UR Research: Open Algorithms, Data and Audits](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/3TATc1wY.jpg)

[UR Research: Open Algorithms, Data and Audits | UR](https://ur.xyz/research?ref=tao.media)

For example, UR’s multi-client routing algorithm keeps a window of network providers available and shifts traffic based on measured transfer performance. Its transfer layer ranks available transports and distributes traffic among them, reducing the share sent through a slow or failing connection.

## Why the Fastest Raw Score Did Not Win

URnetwork [shared results for every scored submission](https://x.com/urnetwork/status/2119878756705640852?ref=tao.media), not just the headline improvement. The [published Round 1 result file](https://api.bringyour.com/competition/v2/rounds/01a1855d-49bd-0250-90db-d0c0885f0e46/result?ref=tao.media) records 69 completed submissions and identifies 14 eligible candidates.

The entry ranked tenth by raw score but first among eligible candidates. Its score was approximately 7,062, against a baseline of approximately 19,556, consistent with the rounded request times in URnetwork’s announcement.

Eligibility depends on six sets of checks: request success, traffic volume, path integrity, matchmaking performance, stability, and resource use. A submission must improve timing while staying within the competition’s operating constraints, rather than winning solely on a lower latency number.

The entry with the lowest raw score illustrates the difference. It scored approximately 5,330 but failed the traffic-volume gate. It generated 2,964 requests, above the permitted maximum of 2,927.4, and received about 3.65 billion bytes, above the allowed ceiling of about 3.60 billion. Its faster timing therefore did not qualify it to win against the baseline workload.

The selected winner passed all six gate categories. Its recorded request success rate was 100%, and its request and byte counts remained inside the required bands. The result file also includes nine evaluation replicates for the entry, alongside statistical significance calculations.

The reward decision remains provisional, with settlement listed as pending at the time of reporting. Integrity review and source approval are both marked approved.

## Each Winner Raises the Next Round’s Baseline

![URnetwork Latency Competition Page](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/11/rLkSbYft.jpg)

[URnetwork Latency Competition Page | Apex by Macrocosmos](https://apex.macrocosmos.ai/competitions/15?ref=tao.media)

The [URnetwork Latency competition](https://apex.macrocosmos.ai/competitions/15?ref=tao.media) runs on Apex, the Subnet 1 platform built by [Macrocosmos](http://macrocosmos.ai/?ref=tao.media). UR’s schedule spans six rounds of one week each, with submissions evaluated against a branch of the protocol’s source code. Winners earn SN1 alpha, the subnet’s token.

The rolling-baseline design makes each round cumulative. Instead of repeatedly measuring entrants against the original code, the next round starts from the previous winner. Later submissions must improve on a system that has already absorbed earlier competition gains.