Potion static embeddings shrink to 700KB

This is a practical deployment win, not just a compression stunt. If the benchmark claims hold up outside the author’s setup, the tiny variants make static embeddings attractive anywhere cold starts, CPU-only inference, or edge deployment matter more than squeezing out the last few quality points.

• –The 700KB micro model is the most interesting piece: it pushes embeddings into browser-extension, WASM, and embedded-device territory.
• –The quality/speed tradeoff looks reasonable for many production workloads, especially if the alternative is a much larger transformer for simple semantic search or routing.
• –The family approach is smart because teams can pick the right point on the size-quality curve instead of overcommitting to one model.
• –The release is also a vote of confidence for model2vec and tokenlearn as a real static-embedding stack, not just a research curiosity.