Using seemingly every model, the predictive text often gives suggestions which are almost completely useless when in the middle of a word.
Smaller models seem to have this issue less often, with only a few issues while using Qwen 3.5 0.8B, many issues while using Qwen 3.5 2B, and a ton of issues while using Gemma 4 E2B. (I have not tested Gemma E4B or Qwen 3.5 4B).
Qwen 2B will often give suggestions for how to finish a word that are a single letter, often which cannot be used to make any English word or any common acronym.
Gemma E2B often gives suggestions which contain random characters (such as $), words that would not create a valid English word, and entirely too many periods.
Using seemingly every model, the predictive text often gives suggestions which are almost completely useless when in the middle of a word.
Smaller models seem to have this issue less often, with only a few issues while using Qwen 3.5 0.8B, many issues while using Qwen 3.5 2B, and a ton of issues while using Gemma 4 E2B. (I have not tested Gemma E4B or Qwen 3.5 4B).
Qwen 2B will often give suggestions for how to finish a word that are a single letter, often which cannot be used to make any English word or any common acronym.
Gemma E2B often gives suggestions which contain random characters (such as $), words that would not create a valid English word, and entirely too many periods.