RunpodReviews of runpod.io

A year ago I had zero background in AI infrastructure — couldn't have told you what a GGUF file was. Claude, Anthropic's AI, walked me through the whole build step by step, explaining the "why" behind every flag rather than just handing me commands. The centerpiece: hybrid CPU/GPU inference running a 35B parameter model on a single RTX 3090 (24GB VRAM) — something that shouldn't fit, but does, using `--override-tensor exps=CPU` to force the MoE expert weights onto system RAM while the rest stays on GPU. We compiled llama.cpp from source with CUDA support (`-DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86` for the 3090's architecture) to get there. On RunPod specifically: Secure Cloud gave me a persistent network volume that survives pod restarts and rebuilds — genuinely useful when I've had to recover from scratch more than once. $0.50/hr for 24GB VRAM, 125GB RAM, and 32 vCPUs is a fair trade for not owning the hardware. Storage is cheap enough to just leave the volume parked between sessions. Also running alongside the main model: a lightweight CPU-only summariser, an SDXL image generation pipeline, and a vision model — all on the same box, all started with one script. Getting all of that talking to each other cleanly took real trial and error (KV cache quantization, context window tuning, flash attention flags), but RunPod never got in the way of any of it. If you're starting from zero and want to actually understand your stack rather than just running someone else's Docker image, Anthropic's Claude paired with RunPod's compute is a combination that works. NOTES: Most of the build work was done with Claude Sonnet 5 thinking, which handled the day-to-day debugging and config well. One genuinely gnarly problem did need a switch to Claude Fable 5 to get unstuck — worth knowing the model choice can matter for the harder edge cases, not just picking whichever is default. Also — once you've got your setup exactly how you want it, remember RunPod's own documentation is explicit that network volumes aren't meant as a permanent storage guarantee. Don't treat a pod/volume as a "set and forget" backup. I set up a proper backup pipeline to Google Cloud Storage and it was genuinely painless — a scheduled sync job, and I've since done a full restore onto a fresh volume from that backup with no real drama. Cheap insurance for something that took real hours to build.

Pretty cool place to launch GPU-based systems for running small-scale pretraining and finetuning experiments.

ive just had so many issues and wasted so much time and money just trying to troubleshoot and get things to work. i understand that the community models i used are a kind of 'use at your own risk situation' but its also happened on official ones, mostly with taking forever to open the pod but never knowing if i just need to wait or restart, go somewhere else, etc. it would be nice to be able to get some credits when things do not work, it would be nice to have a better more streamlined way to make reporting broken template, and the ai assistnat needs to be taught that you should just end a pod if its giving you trouble or doesnt work instead of trying to get YOU to fix it, as if its something that is my problem or responsibility. they either work or they dont, simple as.

The best neocloud. Period. Best service, amazing pricing. And the ability to stop a pod without killing the state is super useful for development. Just missing some mi350Xs for me to really ask who's feet to kiss 😅

I've had an excellent experience using RunPod. The platform stands out for its transparency, efficiency, and reliability. Everything is clearly presented, performance is consistently strong, and the overall user experience has been smooth. I'm very satisfied and would definitely recommend RunPod to anyone looking for dependable GPU cloud services.

I've found their portal easy to se and I like the serverless feature. Had one support incident and it was addressed within minutes in a very friendly way.

runpod one of the best tool

Very useful because it can be operate by AI agents through SSH connection and CLI. I think the configuration input of SSH public key should be placed in a more prominent location and clear instruction should be provided.