I was planning to code a bit on the plane and wanted to use local LLM only to realize that my laptop RAM can only run these two DeepSeek models:

DeepSeek-Coder-V2-Lite
DeepSeek-R1-Distill-Qwen-1.5B

I thought the future was localized LLM with privacy 🥲 

PS: the peak spec for macbook pro max has 128gb but still...

vibe coding @farcaster || prev BS/MS Berkeley EECS

i mean it depends on how you’re using it (chat or autocomplete or fully one shotting code) but models distills have never been worth the effort for me, i’ve had good experience with qwen coder for autocomplete and agent but gemma 3 12b QAT is what i use as my local chat partner

deepseek coder should be decent? also llama 3 needs more RAM unless i quantized down to 4/8 bit lol + context window will be tiny due to all the bits left for input

never used gemma. will take a look 🫡

i know the deepseek distills sounds enticing but for local stuff you really can’t beat the gemma / llama coding fine tunes