(not recommendation, I've not used it .. yet)

Just tried it and honestly it's a terrible experience lacking any sort of intent or reason.

Which is unsurprising in the AI space.

You get a wall of text showing you various random fine-tuned models by random people, and that is basically it.

Actual sane default requirements like "just give me the normal AI labs", "please filter for dense only" and "I want this exact context size at this quant" are not part of the tool, apparently. Neither is "compare these quants for me for the same model".

Or maybe it's just hidden enough that I did not find them before I've stopped caring.

Conway's law is at it again.

____

Edit:

I have since then had qwen3.6 ponder the codebase and think about my complaints.

Seems to require a major data model overhaul to actually fix those, so they're legit. Which I didn't doubt, but nice to have some extra fabricated confirmation after it initially refused and said "nooooo the readme says otherwise nooo hypfer is just a hater noo"

___

Edit 2:

It gets worse the longer I stare at it. This could've been a web calculator.

▲

hypfer 9 days ago | parent | next [-]

Done:

https://github.com/Hypfer/will-it-fit-llama-cpp

https://hypfer.github.io/will-it-fit-llama-cpp/

▲

hparadiz 9 days ago | parent | prev | next [-]

We need benchmarks by engine, cli switch sets, and device with filters by cpu, gpu, and type. And if someone could please aggregate that in a way where people can upload results and just automatically see the best of any model for their device that would be a killer app.

	▲	alfiedotwtf 9 days ago \| parent [-]
		I've wanted to vibe code a tuning app, that pumps data through your CPU-GPU-RAM to try and determine the best parameters for each model, but I think it's just too much work compared to manually running by hand a one-liner and changing things here and there.

▲

dofm 9 days ago | parent | prev [-]

I have found these things to be fully exasperating, to be honest, even though I am seeking information about a pretty "known" machine — a 64GB M1 Max MBP.

(Honestly I think Apple's "AI push" could do worse than just focus on a curated model library, a couple of Apple-standard Gemini distillations, an OS-level model manager and some sort of tweak of their containers system to do what Docker's sbx does. They could demystify a lot of this shit.)