Bento hits small-model serving wall

The model part is almost the easy half now; the hard half is distribution, latency, and unit economics.

• –Whisper handled transcription, so the fine-tune only had to classify and normalize a narrow command schema.
• –Synthetic data plus brutally literal evals made the iteration loop fast, which is why the approach worked at all.
• –The shipping math is ugly: 148 MB for Whisper plus 2.49 GB for the GGUF is a lot to ask users to install for one niche feature.
• –Cloud inference did not really save it either, since always-on L40S-class GPUs can drift into four-figure monthly costs.
• –The piece is a good reminder that packaging and serving are now product design problems, not just ML problems.