One of the most frequent questions one faces while running LLMs locally is: I have xx RAM and yy GPU, Can I run zz LLM model ? I have vibe coded a simple application to help you with just that.
Update: A lot of great feedback for me to improve the app. Thank you all.
Nice concept – but unfortunstely I found it to be incorrect in all of the examples I tried with my Mac.
It'd also need to be much more precise in hardware specs and cover a lot more models and their variants to be actually useful.
Grading the compatibilty is also an absolute requirement – it's rarely an absolute yes or no, but often a question of available GPU memory. There's a lot of other factors too which don't seem to be considered.
> Can I Run DeepSeek R1
> Yes, you can run this model! Your system has sufficient resources (16GB RAM, 12GB VRAM) to run the smaller distilled version (likely 7B parameters or less) of this model.
Last I checked DeepSeek R1 was a 671B model, not a 7B model. Was this site made with AI?
This doesn't mention quantisations. Also, it says I can run R1 with 128GB of ram, but even the 1.58 bit quantisation takes 160GB.
This just isn’t right. It says I can run a 400+ parameter model on my M4 128gb. This is false, even at high quantization.
> One of the most frequent questions one faces while running LLMs locally is: I have xx RAM and yy GPU, Can I run zz LLM model ?
In my experience, LM Studio does a pretty great job of making this a non-issue. Also, whatever heuristics this site is based on are incorrect — I'm running models on a 64GB Mac Studio M1 Max that it claims I can't.
How exactly does the tool check? Not sure it's that useful since simply estimating via the parameter count is a pretty good proxy, then using ollama to dl a model for testing works out pretty nicely.
I think I would like if it also provided benchmarks. The question I have is less can I run this model, but what is the most performant (on some metric) model I can run on my current system?
- When you press the refresh button, it loads data from huggingface.co/api, doing the same request seemingly 122 times within one second or so
- When I select "no dedicated GPU" because mine isn't listed, it'll just answer the same "you need more (V)RAM" for everything I click. It might as well color those models red in the list already, or at minimum show the result without having to click "Check" after selecting everything. The UX flow isn't great
- I have 24GB RAM (8GB fixed soldered, extended with 1x16GB SO-DIMM), but that's not an option to select. Instead of using a dropdown for a number, maybe make it a numeric input field, optionally with a slider like <input type=range min=1 max=128 step=2>, or mention whether to round up or down when one has an in-between value (I presume down? I'm not into this yet, that's why I'm here / why this site sounded useful)
- I'm wondering if this website can just be a table with like three columns (model name, minimum RAM, minimum VRAM). To answer my own question, I tried checking the source code but it's obfuscated with no source map available, so not sure if this suggestion would work
- Edit2: while the tab is open, one CPU core is at 100%. That's impressive, browsers are supposed to not let a page fire code more than once per second when the tab is not in the foreground, and if it were an infinite loop then the page would hang. WTF is this doing? When I break the debugger at a random moment, it's in scheduler.production.min.js according to the comment above the place where it drops me </edit2>.
Edit: thinking about this again...
what if you flip the whole concept?
1. Put in your specs
2. It shows a list of models you can run
The list could be sorted descending by size (presuming that loosely corresponds to best quality, per my lay person understanding). At the bottom, it could show a list of models that the website is aware of but that your hardware can't run
This is where I hope HN has a downvote option. This is not erroneous to the point I want to flag, but the quality is low enough that I want to counteract the upvotes. This is akin to spam in my opinion.
UX whine: Why do I have to click "Check compatibility"? After type and RAM, you instantly know all the models. Just list the compatible ones!
It doesn't have any of my Nvidia GPUs, nor my AMD GPUs in the list and then always tells me I need more VRAM since I can't select a GPU.
128 GB of cpu memory seems to be a bit of a lower upper limit. Maybe it could be increased?
Also, my Mac with 36 GB of memory can't be selected.
Neat, I'd prefer to it just show what models I can run though, rather than saying if I can or cannot run a specific one.
You forgot to define what constitutes “running”. And people have different expectations.
Mhmm.. AMD APU do not have a GPU but can run up to 14B models quite fast
Cool idea. iPhone would be great too! & Function calling tools.
I have ran Deepseek R1 on my PC with 128 gigs ram Effortlessly
Great, but it should have image and video models too.
AI generated app to generate AI generated trash?
I have found the stoplight chart on https://www.canirunthisllm.net/ to be the most useful one of these types of calculators.
even add quantized models
OP, did you even do a simple smoke test of all the options before? None of this works well.
And herein lies the problem with vibe coding - accuracy is wanting.
I can absolutely run models that this site says cannot be run. Shared RAM is a thing - even with limited VRAM, shared RAM can compensate to run larger models. (Slowly, admittedly, but they work.)