I tried out the 8B deepseek and found it pretty underwhelming - the responses were borderline unrelated to the prompts at times. The smallest I had any respectable output with was the 12B model - which I was able to run, at a somewhat usable speed even.