I plan on downloading a Q5 or Q6 version of the 27b for my 3090 once someone puts quants on HF, loading it in LM studio and starting the API server to call it from my scripts based on openai api. Hopefully it's better at code gen than llama 3 8b.