Hacker News

Show HN: Shoehorn – Quantize any model down to run on your machine

69 points by rhgraysonii ago | 17 comments
Working on Mac, Linux, and Windows now. I include a simple GUI to find new models and get things built and set up. It is working quite well across a few models for me. The GitHub README and DESIGN.md files go into detail of the how/why and it's working remarkably well so far. https://github.com/notactuallytreyanastasio/shoehorn

puttycat |next [-]

This is really impressive. Can you say a bit about the underlying process? I'm guessing this is post-training qantization? Isn't PTQ also resource-intensive? (Ie might not work on any machine)

akshay_akula |next |previous [-]

This is interesting. I wonder how it could work with something like https://github.com/JustVugg/colibri.

hmokiguess |next |previous [-]

Reminds me of https://github.com/AlexsJones/llmfit

rhgraysonii |root |parent [-]

LLMFit tells you what can run on something. I built something quite similar to their search into Shoehorn now.

sscarduzio |next |previous [-]

The project name is perfect!

jedbrooke |next |previous [-]

I gotta laugh at some of the models it suggests, for example:

> AnkitAI/Parable-Qwen3-4B-Claude-Fable-5-GGUF

you’re telling me you managed to fit Fable 5 into just 4B?

chompychop |root |parent [-]

I gotta laugh at your thought process: knowing Fable 5 is a large frontier model, you're telling me that the first thing that came to your mind on seeing that model name is that it's a quantized version of Fable? As opposed to a distillation/fine-tuning on Fable responses?

metalliqaz |root |parent |next [-]

Well to be fair here... the title of this post doesn't mention fine tuning, it mentions quantization.

unrented7977 |root |parent |previous [-]

Don't make fun of people you think are ignorant, it's a pretty shitty look

chompychop |root |parent [-]

Well, then don't get all snarky and dismissive of things you might not be knowledgeable about ("you" here referring to OP).

mbuchel-hn |next |previous [-]

does this work similar to airllm? i am wondering how it would handle something like quantizing kimi k3 on a budget of 8 gbs, or is that something you are not attempting to solve yet?

rhgraysonii |root |parent [-]

Yes that is exactly what this does.

kennywinker |root |parent [-]

Could you explain what happens when you try to shoehorn a 2.4T parameter model into a 24gb m4 mac?

metalliqaz |root |parent |next [-]

extreme divergence would be my guess

akshay_akula |root |parent |previous [-]

Wondering the same thing but for 48gb M5 Max.

jaylane |next |previous [-]

tried it out but based on the model sizing result i got i got an insufficient memory error when the server started running

rhgraysonii |root |parent [-]

If you could post an issue if you still have the error around that would be awesome.

kelvo_ran |previous [-]

[dead]