Incredibly slow download speeds | Runpod | Page 1

white crypt Feb 13, 2026, 7:25 PM

#

Hi there, I've rented a pod for ~6$/h and spend a good bunch of time just to run system upgrades and downloading a package that is a few hundred megabytes. I ran a speed test and while it says 9.85 MBit/s as download speed, it really doesn't feel like it's downloading with anything near that (while 9.85 MBit/s is also slow compared to all other pods I've rented so far). I'll attach an image of the speed test in a second.

signal zenithBOT Feb 13, 2026, 7:25 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

fierce micaBOT Feb 13, 2026, 7:25 PM

#

white crypt Feb 13, 2026, 7:25 PM

#

#

I've decided to terminated by the way, if I start downloading data, I'll pay probably 20$ just for downloading my data...

eternal vapor Feb 14, 2026, 10:08 AM

#

Speedtest-cli is not an accurate tool

#

All hosters hate it because it constantly underreports

#

Which region was it and was it secure cloud or community cloud?

#

I can spin one up and check

white crypt Feb 14, 2026, 10:09 AM

#

It was secure cloud in iceland

eternal vapor Feb 14, 2026, 10:09 AM

#

Do you remember the number?

white crypt Feb 14, 2026, 10:10 AM

#

Can I check in my pod history or something? I can't say it exactly, but it was probably 1

eternal vapor Feb 14, 2026, 10:10 AM

#

Nah but ill try 1 then

white crypt Feb 14, 2026, 10:11 AM

#

Regardless of the speedtest results, it took at least 10 minutes to download 2% of the ollama tar. On the newly spun up server (other location), in 10 minutes I had the system updated and ollama was serving already...

#

It's not an issue if you pay 0.6$/h but with 5$ an hour it's kind of frustrating

eternal vapor Feb 14, 2026, 10:12 AM

#

#

Use https://koboldai.org/koboldcpp its far more optimized for runpod

#

You wont have to deal with the dependency stuff just edit the template with the GGUF link and context size you want and there you go

#

Can confirm the usual 400mb/s download speed from huggingface in iceland 1

#

Actual speed not speedtest speeds

white crypt Feb 14, 2026, 10:21 AM

#

Again, I don't really care much about the speedtest result as long as the things I need to download to get working take ages. I understand that it's difficult to debug in hindsight but keeping the server alive for half a day just to debug is too expensive.

#

I probably shouldn't even come to the discord to report it since I decided to terminate the pod rather quickly

eternal vapor Feb 14, 2026, 10:22 AM

#

Would be hard to see in hindsight but I can confirm iceland with koboldcpp downloads fast

white crypt Feb 14, 2026, 10:22 AM

#

Also, a quick speed test on the second pod I spun up revealed download speeds of a factor 10

#

Can you try to apt update & upgrade?

eternal vapor Feb 14, 2026, 10:23 AM

#

It did at the beginning

white crypt Feb 14, 2026, 10:23 AM

#

Did it go through perfectly fine? That would be shocking haha

eternal vapor Feb 14, 2026, 10:23 AM

#

Or apt install for ssh at least

#

I do have a template that does really comprehensive apt stuff if you need a more thorough check on my end

#

But you could avoid the hassle by not using ollama xD

white crypt Feb 14, 2026, 10:25 AM

#

I needed to use a language model that doesn't fit on my own workstation and again, I don't think ollama was the culprit since the apt upgrade took way longer than usual as well

eternal vapor Feb 14, 2026, 10:25 AM

#

white crypt I needed to use a language model that doesn't fit on my own workstation and agai...

If you use that llm with koboldcpp your up and running so much easier

#

Which llm is it?

white crypt Feb 14, 2026, 10:26 AM

#

Discussing things in retrospect doesn't seem to help since there are no logs saved or something. I wanted to try the latest deepseek model.

eternal vapor Feb 14, 2026, 10:27 AM

#

Thats very big what gpu did you rent,

#

Because when you said ollama I didnt expect a h200

white crypt Feb 14, 2026, 10:28 AM

#

I set up the VRAM to over 100 GB and rented 5x5090s to see if that'll work. You can imagine that I didn't go ahead and download the 400 GB model and spun up qwen3-coder-next on the second pod I've set up which honestly didn't help much with my problem.

eternal vapor Feb 14, 2026, 10:31 AM

#

5x5090 is expensive for that model

#

Not that much vram either

#

Ill match the budget but on the mi30px

#

Took me a moment but deploying it on 2xMi300x. That wont be a full offload so it wont be as fast as it can be. 500gb of vram is needed for it to be a full offload

#

But it price matches the 5x5090

#

Had to relaunch since its kinda tricky on my phone

#

Lets see how long this takes

#

While thats downloading I spun up qwen3 coder

#

That model is so new its newer than the engine so I hope it works

eternal vapor Feb 14, 2026, 10:50 AM

#

white crypt I set up the VRAM to over 100 GB and rented 5x5090s to see if that'll work. You ...

https://mylrxp1crcd070-5001.proxy.runpod.net/ qwen3 coder took 3 minutes to boot

KoboldAI Lite

KoboldAI Lite - A powerful tool for interacting with AI directly in your browser. Chat with AI assistants, roleplay, write stories and play interactive text adventure games.

#

Deepseek took 20 https://jjmt88lakhc105-5001.proxy.runpod.net/

KoboldAI Lite

KoboldAI Lite - A powerful tool for interacting with AI directly in your browser. Chat with AI assistants, roleplay, write stories and play interactive text adventure games.

#

But 20 for a model that large is good haha

white crypt Feb 14, 2026, 1:30 PM

#

I would've been totally fine with 20 minutes. I've spent 20 minutes for the download before terminating.

#

Thanks for trying out though!

eternal vapor Feb 14, 2026, 7:14 PM

#

20 minutes for a 400gb model file is very impressive haha

white crypt Feb 16, 2026, 1:25 PM

#

I've canceled after 20 minutes, it was closer to stalling than the 400 GB boing downloaded somewhere soon

#Incredibly slow download speeds