Read throughput, latency, jitter and loss separately
A result describes the connection from your device to the test destination under those conditions. A different destination, Wi-Fi, node or time can change it. Test platforms also differ, so one platform's score does not describe every app.
Record the metrics below alongside the conditions and raw results, then compare like-for-like runs. Do not infer long-term bandwidth from the best single result.
| Metric | What it describes | What it does not prove |
|---|---|---|
| Download / upload throughput | Transfer rate on the tested path | Guaranteed bandwidth for every destination or time |
| Idle / loaded latency | Round-trip delay without or during transfers | Model generation speed |
| Jitter | Variation between latency measurements | Average throughput |
| Packet loss / request failures | Loss or failures in that test method | The failure cause of every protocol |
Change one condition at a time
Confirm the subscription is valid and a node is selected, then record System Proxy, mobile VPN or TUN mode. Keep your configuration and compare one node or protocol at a time, rather than changing the client, DNS and Wi-Fi together.
Use the same destination and pause large downloads or updates. As a practical method, repeat each group at least three times, compare its median and failures, then run another group at your usual time. This does not establish a long-term reliability guarantee.
Cloudflare Speed Test is one available tool, measuring its own test network rather than proving the route to an AI platform. Speed tests transfer data; uploads and downloads count toward HKUVPN's allowance.
- Record the device, client version, local network, node and protocol.
- Record local date, time, time zone and test destination.
- Keep every result and failure count before summarizing the median, rather than selecting the fastest run.
A TCP test cannot validate a failing HY2 UDP connection
HY2 (Hysteria2) uses UDP. A successful TCP port connection does not validate its UDP path, authentication, certificate or obfuscation. Use real client logs and browsing tests; a timeout can have several causes, as covered in official troubleshooting.
Try the subscription's Trojan option on the same network as a comparison. If Trojan works while HY2 times out, record each node, time and log before contacting support. Retain the subscription's certificate and obfuscation configuration.
AI first-response time depends on more than the network
Waiting for an AI response includes connection setup, provider queuing, model processing and returned content. Test ordinary HTTPS first, then inspect the AI request's status and timing. Lower latency does not guarantee faster inference or better answers.
If the browser works while a terminal or SDK times out, first check its proxy. For responses such as 401, 403 or 429, follow the provider's authentication, permission or rate-limit guidance rather than treating every error as a VPN speed issue.
Give support reproducible results
Provide the failing step, device and client version, network type, node and protocol, test time, destination and error text. A working and failing node compared under the same conditions makes the next check easier to identify.
Redact full subscription URLs, API keys, passwords, verification codes and session credentials from screenshots and logs. These are methods for your own tests, not measured HKUVPN performance results.