Is this your problem? A quick decision table
Bufferbloat is a specific failure: latency rises sharply while the connection is busy, because data waits in an oversized queue somewhere along the path. The usual symptom is a call that is fine until someone starts an upload, a backup or a large download.
| What you see | Likely next step | Skip SQM when |
|---|---|---|
| Calls fail only while another device uploads or downloads | Run a loaded-latency test, then consider SQM | Loaded latency stays close to idle latency |
| Calls fail on one laptop, even when nothing else runs | Check Wi-Fi signal and placement first | The wired test from the same desk is clean |
| Upload total from calls is near the plan's upload rate | Reschedule backups or change plans | SQM cannot create capacity the line does not have |
| Router CPU is already busy with VPN or other features | Check model support and expect lower throughput | The router is not on GL.iNet's SQM support list |
Measure latency under load, not just speed
Cloudflare describes its speed test as recording idle latency as soon as the page loads, then sending small empty requests at intervals during the download and upload phases to measure loaded latency. It also reports idle and loaded jitter, which it defines as the average variation between consecutive round-trip measurements, along with packet loss. The gap between idle and loaded latency is the useful number for this problem. [1]
OpenWrt's SQM guide suggests testing with Waveform, Speedtest, LibreQoS or Cloudflare while the network is otherwise idle, and says using more than one test helps rule out a bottleneck in one test. It recommends Waveform when tuning settings. GL.iNet's SQM blog also starts with Waveform's online bufferbloat test. [2][5]
Use one wired computer for the baseline so Wi-Fi does not muddy the result. Run the test at a quiet time, then again at the time calls usually fail. Write down idle latency, loaded latency during download, loaded latency during upload, and the throughput figures. Keep this record before you change any router setting.
What SQM does, and its costs
OpenWrt describes SQM as an integrated system that performs per-packet and per-flow network scheduling, active queue management, traffic shaping and prioritization. In practice it limits the router's rate to slightly below the line rate. Packets then queue in the router, where a modern queue discipline can manage them, and not in the modem or ISP equipment. Cake is OpenWrt's preferred option for bufferbloat, and fq_codel is described as a faster but less comprehensive alternative. [2]
There are costs. OpenWrt's starting point is 90 percent of the measured speeds, so you give up some peak throughput. Its guide says SQM runs on the CPU and slower devices may not keep up with a full connection. It also says hardware flow offloading must be disabled, while software flow offloading can stay on. On a fast line with a modest router, the shaper itself can become the bottleneck. [2]
Link-layer settings matter on some connection types. OpenWrt's guide asks you to choose your link type and overhead. If you are unsure, it suggests overestimating with Ethernet and an overhead of 44 and an mpu of 96. [2]
SQM and QoS on GL.iNet firmware 4
GL.iNet's firmware 4 docs describe built-in SQM, introduced in firmware 4.9, that applies when the router acts as the gateway. You set upload and download limits; on 4.11 and later you can enter them or run a built-in speed test. You choose cake, which the docs recommend, or fq_codel. Cake Autorate is an option on 4.11 and later for links whose bandwidth changes, such as cellular. The docs say it is not needed on stable links. Enabling SQM automatically turns off Network Acceleration, and SQM and QoS cannot run together. Some models lack the memory for SQM; Flint 2 and Beryl AX are on the supported list. [3]
GL.iNet's separate QoS feature, also introduced in 4.9, offers Device Priority, which gives chosen clients priority when the WAN is congested. It also turns off Network Acceleration when enabled. Its supported list includes Flint 2 and Beryl AX and excludes models such as Opal. [4]
GL.iNet's blog takes a narrower view. It is written for connections where limited upload causes congestion. It describes installing SQM packages through the LuCI interface, setting download to 0 so that only upload is shaped, and setting upload to about 90 percent of the measured rate. It advises against SQM on fast connections that are not congested. Where the blog and the firmware docs differ, follow the docs for your firmware version and treat the blog as an upload-only starting point. [5]
On 13 October 2026 GL.iNet's store listed Flint 2 at $169.99 and Beryl AX at $98.99; confirm checkout. This publication has not run SQM on either model. Do not buy a router for SQM before your current router's support is checked.
Worked example: a fictional upload-limited line
Fictional household: a cable plan with modest upload. Calls break up whenever a laptop backup runs. A wired loaded-latency test shows idle latency in the low tens of milliseconds, rising to several hundred during the upload phase. Download-phase latency barely moves.
The plan follows GL.iNet's blog for an upload-limited line, which uses the LuCI SQM packages; on the built-in firmware SQM page, use the options your firmware version offers. Record the current settings and firmware version. Turn on SQM in LuCI with download set to 0, which the blog uses to leave download unshaped, and upload at about 90 percent of the measured rate. Choose cake. Rerun the same test at the same time of day with the backup running, and record whether upload-phase latency drops and how much upload throughput was given up. If latency improves, raise upload gradually until it starts to climb again, then back off. That is OpenWrt's tuning loop. If nothing improves, turn SQM off and note that the queue is somewhere SQM cannot reach. [2][5]
These are planning steps with illustrative numbers. They are not a measured result.
Pitfalls and who should leave SQM off
Write the router's original settings down before changing anything, and keep a one-line rollback note. The firmware change plan guide covers backups and recovery if a setting change locks you out.
- Leave SQM off if loaded latency is already close to idle latency; you would give up throughput for nothing. [5]
- Do not run QoS and SQM together on GL.iNet firmware 4; the docs say they cannot be enabled at the same time. [3]
- Expect Network Acceleration to switch off, and check CPU headroom if you also run a VPN on the router. [2][3]
- Do not set limits from the plan's advertised speed; measure first, then start below the measured rate.
- Do not judge a change from one test; repeat at the time calls usually fail.
- On cellular or other variable links, read GL.iNet's Cake Autorate notes; fixed limits may be too high or too low through the day. [3]
- Microsoft also recommends prioritizing media over Wi-Fi and keeping Teams traffic out of a VPN tunnel where possible; check those before blaming the WAN. [6]
Questions before you decide
Will SQM make my internet faster?
No. SQM usually lowers peak throughput a little, because the shaper is set below the line rate. The aim is lower and steadier latency when the line is busy. GL.iNet's own blog suggests leaving SQM off on fast connections that are not congested in normal use.
Can I run QoS and SQM together on a GL.iNet router?
Not on firmware 4. GL.iNet's docs state that QoS and SQM cannot be enabled at the same time. Both turn off Network Acceleration. Pick one, test it, and keep a note of the original settings.
Which GL.iNet models support SQM?
GL.iNet's SQM page lists supported models and says the feature arrived in firmware 4.9. Some models lack the memory for it. Flint 2 (GL-MT6000) and Beryl AX (GL-MT3000) appear on the list. Check your exact model and firmware before planning around it.
Sources & verification
Product details and prices can change. Check the linked provider before buying.
- Cloudflare blog: How does Cloudflare's speed test really work? Accessed 13 Oct 2026
- OpenWrt: SQM (Smart Queue Management) Accessed 13 Oct 2026
- GL.iNet Router Docs 4: SQM Accessed 13 Oct 2026
- GL.iNet Router Docs 4: QoS Accessed 13 Oct 2026
- GL.iNet blog: How to reduce bufferbloat with SQM on GL.iNet routers Accessed 13 Oct 2026
- Microsoft Learn: Prepare your organization's network for Teams Accessed 13 Oct 2026
Sources link directly to providers. Product buttons may use separately labeled affiliate links. Read our disclosure.