Post

My e-ink tablet's handwriting recognition has a 27% error rate. So I benchmarked local VLMs against cloud models: Claude Opus 4.5: 3% error, 6 seconds qwen3-vl:8b (local): 5% error, 74 seconds A 2% accuracy hit to keep my journal entirely on-device? I'll take that trade. smus.com/notes/2025/l...