# @jmbollenbacher — 2026-08-27

♥20 ↻1 · https://x.com/jmbollenbacher/status/2092803393299399034

I still wanna know how Luna is so token efficient compared to its peers (dsv4-flash, glm5.3-flash, etc.).

It uses like 40% of the tokens for comparable results.

How do they do it? It's not a way bigger model. It's some kinda posttraining trick. How?

tags: author:jmbollenbacher, kind:tweet, on:glm-5-3-flash, year:2026
cited on: _dossiers/glm-5-3-flash.md, glm-5-3-flash