Which Ollama model would be the best?
I wanted to get the Deepseek R1 8B model, but ChatGPT says my PC will cry if I install that.
Specs:
Can I run any AI at all (locally)?
Your laptop won’t just cry – it’ll beg for mercy before bursting into flames… 
~5GB of free disk space isn’t enough even for just downloading Deepseek R1 8b.
As for the CPU and GPU specs, the LLM will have to be handled all by the CPU, which the Ryzen 3 3250U will struggle with.
But even with a much more powerful CPU, 6GB of RAM isn’t enough to hold Deepseek R1 8b (the “8B” is short for “8 billion tokens”) in memory while it runs without thrashing swap space.
The “1.5b” Deepseek R1 model would fit within your laptop’s hardware specs, but it’s not going to be fast.
Thanks.
Regarding the disk space thingy, the website lists the size for the 8B model as 5.2 GB. How’s it not enough?
For context, I have several gigs of free space. I just said I wanted a model around 5GB big, not “significantly” more, e.g., not models of size 8 GB, 10GB…
Is the installed size of the model bigger than the listed size of the model on the website?
Yes, the 8b model is 5.2GB, but combined with filesystem overhead, there’s a chance it won’t fit within the ~5GB available on your NVMe SSD.
I mean, I have 60 GB of free space on the SSD…
The ~5GB was constraint for the size of the model I’d install.
Because that 60 GB is precious for me.
Aside from storage you need far more memory to run any vaguely usable model, unfortunately.
Ollama itself (not the model) says it would need 8 GiB to install on my machine:
$ dnf install ollama --assumeno
Updating and loading repositories:
Fedora 43 - x86_64 - Updates 100% | 17.7 KiB/s | 16.8 KiB | 00m01s
Repositories loaded.
Package Arch Version Repository Size
Installing:
ollama x86_64 0.9.4-4.fc43 fedora 772.8 MiB
Installing dependencies:
hipblas x86_64 6.4.1-4.fc43 fedora 1.1 MiB
hipcc x86_64 19-14.rocm6.4.2.fc43 fedora 652.9 KiB
rocblas x86_64 6.4.4-1.fc43 updates 4.2 GiB
rocm-clang x86_64 19-14.rocm6.4.2.fc43 fedora 70.2 MiB
rocm-clang-devel x86_64 19-14.rocm6.4.2.fc43 fedora 23.3 MiB
rocm-clang-libs x86_64 19-14.rocm6.4.2.fc43 fedora 98.4 MiB
rocm-clang-runtime-devel x86_64 19-14.rocm6.4.2.fc43 fedora 7.8 MiB
rocm-comgr x86_64 19-14.rocm6.4.2.fc43 fedora 123.9 MiB
rocm-device-libs x86_64 19-14.rocm6.4.2.fc43 fedora 3.2 MiB
rocm-hip x86_64 6.4.2-2.fc43 fedora 24.9 MiB
rocm-libc++ x86_64 19-14.rocm6.4.2.fc43 fedora 1.2 MiB
rocm-libc++-devel x86_64 19-14.rocm6.4.2.fc43 fedora 7.5 MiB
rocm-lld x86_64 19-14.rocm6.4.2.fc43 fedora 5.7 MiB
rocm-llvm x86_64 19-14.rocm6.4.2.fc43 fedora 48.5 MiB
rocm-llvm-devel x86_64 19-14.rocm6.4.2.fc43 fedora 25.3 MiB
rocm-llvm-filesystem x86_64 19-14.rocm6.4.2.fc43 fedora 0.0 B
rocm-llvm-libs x86_64 19-14.rocm6.4.2.fc43 fedora 84.8 MiB
rocm-llvm-static x86_64 19-14.rocm6.4.2.fc43 fedora 1.8 GiB
rocm-runtime x86_64 6.4.2-3.fc43 updates 3.1 MiB
rocm-runtime-devel x86_64 6.4.2-3.fc43 updates 571.4 KiB
rocsolver x86_64 6.4.4-1.fc43 updates 987.9 MiB
Transaction Summary:
Installing: 22 packages
Total size of inbound packages is 2 GiB. Need to download 2 GiB.
After this operation, 8 GiB extra will be used (install 8 GiB, remove 0 B).
I don’t know if all those ROCm dependencies are strictly necessary, but you probably need at least some of them to take advantage of your iGPU.
Welp, that’s a real blocker 
I can’t give 15 gigs to an AI for basic chats… chatgpt.com seems to be the only way out it seems 
Ah, 60GB of free space is fine then. The 8b model might not load depending on what the swap setup is.
For comparison, I’ve been running Google DeepMind’s 7b “Gemma” model on an AMD EPYC CPU (no GPU assist) with ~16x the multi-core performance of the Ryzen 3 3250U. It’s usable, but takes 6-10 seconds to answer “What is Linux?”.
Maybe Google Colab is an option if you want to run the model directly, except on a cloud computer rather than a local server.
I haven’t used it myself for quite a while,but a quick search gives some howto guides for running Ollama on it.
For reference, Ollama’s GPU support matrix: https://docs.ollama.com/gpu
Try it; I was playing PCVR Beat Saber yesterday on a office laptop with Intel UHD 630 
I run a lot of AI models locally - doing everything from chatting with LLMs to producing images and music. My main rig for AI use has an i7 CPU, 32 GB RAM, and a 12 GB RTX 3000 series GPU. That makes everything run fast and smoothly.
My main workstation desktop has been a trusty old Optiplex with 16 GB of RAM and a modern i7 CPU. That latter computer will run AI models locally, but it’s not hugely fast. I don’t normally run anything over a 4B model with it. For your setup, I definitely would aim for an LLM that is listed as 1.5B (or no more than 3B) range. Your CPU will produce an answer very slowly, though I’m not sure you have enough RAM to run both your OS and an LLM.
If you do get it running, it helps if you tell it something like, “Limit your answer to 200 words” so that it gives you an answer that is produced more quickly.
For those who are new to running local LLMs, LM Studio is a bit more user friendly than ollama.