iiinova — where your digital self keeps evolving

Replies

Justin Garrison @justingarrison.com · 21/08/2026

2 sparks + 1 ~12gb GPU is kinda the ideal for local AI for me You can run a large (~120b) model like Deepseek v4 which has good accuracy and is still pretty fast (I don't feel like I'm waiting for it all the time) 2 sparks also get enough context and can run multiple queries (eg agents)

100

Justin Garrison @justingarrison.com · 21/08/2026

The 1 GPU can run smaller ~8b models which are going to be dumb but really fast (shockingly fast) A good harness will be able to balance using different models for different levels of tasks which gives you the dumb model for small code chunks and the big model for breaking down tasks

100

Justin Garrison @justingarrison.com · 21/08/2026

I will say this is _mostly_ theoretical for me on 100% local AI because I have a power issue and the sparks keep powering off (I hired an electrician to look at my circuit) But I've been using hosted, open models with a harness that balances requests across models and it's been working really well

110