Offline ai and custom ai keys

[sandyy-cpu]

I was testing offline ai (with model of qwen2.5-0.5b-instruct-q4k_m.gguf) on my phone having Snapdragon 8s Gen 3 processor and 2.5 usable RAM with no app running in the background.

Problem I faced:-

1. Too much latency.
2. Primary proofread button is not working at all. Meaning the output is either gibberish or completely random text and sometimes no output.
3. When I tried custom ai key with my custom prompt it pasted the same paragraph many times in the output. It specifically happens when copied text is pasted one or more than one from the clipboard and the key is tried. I tried a shortened one line prompt to reduce tokens for best latency, even after doing this there was no change in latency.
4. The custom ai key button (key 1) also acts as a paste button instead of triggering the ai action.
5. Several other problems I faced like buttons on toolbar that is use for making cursor jump to start or last, this button too is not working best as it should be.
6. Any model recommendations that work best?

My use case as of now- only proofread (like fixing typos capitalization etc)

[LeanBitLab]

I was testing offline ai (with model of qwen2.5-0.5b-instruct-q4k_m.gguf) on my phone having Snapdragon 8s Gen 3 processor and 2.5 usable RAM with no app running in the background.

Problem I faced:-

1. Too much latency.
2. Primary proofread button is not working at all. Meaning the output is either gibberish or completely random text and sometimes no output.
3. When I tried custom ai key with my custom prompt it pasted the same paragraph many times in the output. It specifically happens when copied text is pasted one or more than one from the clipboard and the key is tried. I tried a shortened one line prompt to reduce tokens for best latency, even after doing this there was no change in latency.
4. The custom ai key button (key 1) also acts as a paste button instead of triggering the ai action.
5. Several other problems I faced like buttons on toolbar that is use for making cursor jump to start or last, this button too is not working best as it should be.
6. Any model recommendations that work best?

My use case as of now- only proofread (like fixing typos capitalization etc)

any model less than 1b will always output gibberish, its not a bug but the limit of the model, you can discuss the models in here #16 also try Qwen2.5-1.5B Q4_K_M , your device is capable