Tonight a 26B model at ~20GB became a lot of people's local default. Capability stopped being the constraint - memory bandwidth did, and the runtimes still lag the weights. What are you running locally? #AI #LocalLLM