# what model would you recommend for cpu only on windows 48gb ram — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/72264224a35ff238

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-18 17:43 ingested 2026-11-19 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xzlkre/what_model_would_you_recommend_for_cpu_only_on/)

## retained text



```
OK. so I have a 5060ti with i use for various DiT inference engines like comfyui and audio.cpp. i have an 8845hs cpu with 64gb DDR5 but as this is my main working pc some of the ram is used on browser and other running apps.

Now what I want is a model to run basic stuff like create and run a script to run a folder of audio files through audio.cpp stem separation and transcribing or write a bunch of prompts for klein and generate on my comfyui. I normally use opencode builtin and openrouter free api for this kind of stuff but occasionally when credit runs out i need an alternative.

I have llama.cpp installed so with the right flag it should restrict itself to ram but which model should I go for?

  
Thank you all for the replies - too many to reply individually to - I got gemma-4-26B-A4B-it-qat-UD-Q4\_K\_XL.gguf using the igpu and ram and getting on llama.cpp around 24t/s which is usable for quick executions (for heavier coding I have basalt qfn and coder but these obviously fight for vram with my dit inference engine ) 
```



document fetched 2026-11-19 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzlkre/what_model_would_you_recommend_for_cpu_only_on
