It can take around 10 mins for the file download to even start and when it does
It says it is going to take 11 days to download (I have 500MB internet) and then it fails soon afterwards.
I thought it might be my VPN so turned it off but it still fails. Tried different bowsers, still no, I see no way I can get hold of these model files.
In my environment (Firefox), the speed seems to be over 100 MB/s. I wonder if there is some kind of restriction or if the bandwidth is being throttled somewhere in the network…
Similar to @John6666 , in my own testing (on both Brave and MS Edge) I can download without issue. Web requests are fully supported from our backend, so I’m wondering if the issue lies somewhere else.
Hi, yes I think I have worked out that it is my VPN causing issues (NORD VPN), but VPN’s are a necessity now in the UK because many AI sites I use (like Civtia.com) are blocked completely without one.
The strange thing is that just turning the VPN off is not enough to fix the issue I have to restart my whole PC and then I can download from xet again until I next have to turn my VPN on, it is very inconvenient to keep having to reboot when I have training and generation going on locally.
Your VPN is leaving behind DNS/routes that break Xet-backed HF downloads. You don’t need to reboot. Do one of the following fast fixes.
Fastest stable workarounds
Split-tunnel a “download browser.” In NordVPN on Windows, enable Split Tunneling and exclude one browser from the VPN. Use that browser only for HF downloads. Keep your other traffic on the VPN. (support.nordvpn.com)
Disable Threat Protection Pro for the test. It inspects DNS/URLs and can block or slow downloads. If downloads resume, keep it off during HF downloads or add allow-lists. (NordVPN)
Allow-list Xet + HF endpoints in any filter (Nord Threat Protection, corporate gateway, local firewall):
# disable Nord's TAP/TUN, then re-enable your main adapter
Get-NetAdapter | ? {$_.Name -match "TAP|Nord"} | Disable-NetAdapter -Confirm:$false
Enable-NetAdapter -Name "Ethernet" -Confirm:$false # or "Wi-Fi"
This forces route/DNS refresh without a full reboot. (Microsoft Learn)
Reduce interference from NordVPN
Switch protocol. Try OpenVPN (UDP/TCP) if NordLynx misbehaves, or vice-versa. (support.nordvpn.com)
Fix MTU with WireGuard/NordLynx. Many see success near 1380. Test and set MTU:
Find max unfragmented size: ping 1.1.1.1 -f -l 1472 and decrease until it succeeds; MTU = value + 28.
Set MTU: netsh interface ipv4 set subinterface "Ethernet" mtu=1450 store=persistent (name per netsh interface ipv4 show subinterfaces). (Reddit)
Use non-Nord DNS while disconnected or for the excluded browser. Set 1.1.1.1/1.0.0.1 (or 8.8.8.8/8.8.4.4) at the adapter, or in Nord’s Custom DNS. (Cloudflare Docs)
Consider the NordVPN browser extension for the “download browser” instead of the desktop app; it proxies only the browser’s traffic and avoids system-wide hooks. (support.nordvpn.com)
Browser-specific levers
Fully quit Brave/Edge so they release network state. Disable background apps to ensure a clean restart after toggling VPN:
Edge: edge://settings/system → turn off Startup boost and Continue running background extensions and apps… (Microsoft)
Mitigate WebRTC quirks if your VPN policy cares about IP leaks. Brave/Edge can limit WebRTC; this also changes connection paths. (Avoid the Hack (avoidthehack!))
Prefer CLI for Xet repos
Downloads are often more reliable via the HF tooling, and you can throttle concurrency.
Ensure the same allow-list (above) if filtering is in place. (Hugging Face)
Why this happens
HF moved many large repos to Xet. Xet adds new domains and a CAS flow that some VPN DNS/filters break. When you disconnect NordVPN, Windows can keep Nord’s DNS/routes until you reset the stack, so browsers continue failing until a reboot. The allow-list and resets address that path. (Hugging Face Forums)
