I built a page that tells you what AI model your laptop can run (localsotabenchmark.engineersf.dev)

and no I'm not a software develpoer - just smacked it together with minimax for personal needs.

Thought it was nifty so shared. Would love to see a real CS person's version of this and/or best prompt to build something like this accurately and keep it up to date as models advance and SOTA shifts.

One word: wrong.

Great idea! You're 95% of the way there and the rest is a couple of UI tweaks.

1. I thought the table showing a bunch of models after my laptop specs would be models I could run on my laptop.

    Model Tier Example Models Params RAM Needed Your Status
    Top SOTA GPT-4o, Claude 3.5 Sonnet, Gemini Ultra 1T+ Datacenter —
    Large Models Llama 3 70B, Mistral Large, Command R+ 70B 64 GB  Usable
    Mid Models Llama 3 8B, Mistral 7B, Phi-3 Medium 7-13B 16 GB  Usable
    Small Models Phi-3 Mini, Llama 3 2B, Gemma 2B 1-3B 4 GB  Your Tier

'Your tier' is too subtle. I would make 'Your tier' highlight the ENTIRE row in some bright color to make it more obvious, and maybe fade the other rows slightly.

2. I don't know what 'usable' means. How can these models be status:usable if they are not 'my tier'? I deduce 'usable' doesn't mean usable on my laptop but I don't know what 'usable' does mean.

3. Give me links to set up the best model I can.

For some reason it is not able to detect 64GB RAM on my Fedora Linux mini PC. It shows only 8GB.

That's not very accurate. It sees 8 GB RAM instead of 64. And it detects GTX 9 GPU, though I have an RTX30 one.

(Edit: now it came to my mind that Firefox has fingerprinting protection by default. But even after whitelisting this website, the result is the same)

I find this to be pretty accurate: https://www.canirun.ai/

Identified every component of my system incorrectly. W11.

Interestingly wrong; probably has to do with the fingerprinting and whatnot.

I do appreciate that the "canirun" one

1) was also wrong 2) but let me pick my real card from the dropdown with results that seemed much more accurate.

Genuine question - is there literally any factor besides RAM and GPU (with VRAM info?)

auto-detect failed to correctly detect 128GB ram vs 32GB ram on MacBook Pro M5

Reports android tablet as a 'linux PC'.

Firefox on macOS. site reported M1 and 16GB. Mac is M4 + 24GB

It's too conservative, on my Air M4 I run gemma-4-12B-it-qat-UD-Q4_K_XL fine with llama.cpp

The app states "Small Models (1-3B)".

What about my desktop? Will it detect what models a desktop can run?

I recommend llmfit. It shows you exactly what models you can run and at what speeds

https://github.com/AlexsJones/llmfit

Be nice if I could correct the specs when detected incorrectly, or check the specs of a different machine without needing to physically pull this up on each machine.

Model choices all seem quite out of date. Phi-3, GPT-4o? Claude 3.5? And I thought Gemini Ultra was a subscription plan, not a model.

Ive got a tack server collecting dust.

How do I plug in a specific hardware?