Running local models is good now: Vicki Boykis's hands-on take
Running local models is good now
Vicki Boykis has been running local models on an M2 Mac with 64GB RAM for over a year, and now finds them genuinely useful. Gemma 4 26B and the newer 12B qat variant let her do agentic coding locally at roughly 75% of frontier-model accuracy. She uses Pi as the agent harness and LM Studio for inference, all inside a Docker container with restricted permissions. The post doesn't give token speeds or latency numbers, but notes the KV cache can eat all 64GB of RAM.
Why it matters: Vicki Boykis ran local models on a 64GB M2 Mac for over a year, and recorded that Gemma 4's 26B and 12B qat versions reach about 75% of top closed models' performance on agentic coding.