More meaning. Less context.
Reduce the amount of irrelevant material between intent and useful understanding.
On Supabase, answering five real questions took between 13.9 KB and 148.7 KB of text, against the 30.5 MB of source behind them. The published model of that same scope is 1.57 MB — 19.4 times smaller than the code it describes. In the head-to-head run, GitMir answered from 8.4K tokens of context per question where the best alternative used 27.4K.
Open a finished public model and read what a single question was answered from.
Compare that against the source an assistant without a model would have had to search instead.
Point your own agent at the model over MCP and watch what it spends per question.
Build a model of your own repository and check the same two numbers on it.
Measured on one repository at one version — Supabase's apps/studio and packages/pg-meta, on 9 September 2026. "19x" is not a general claim about all codebases. Two different numbers are reported and must not be merged into one: how much text a single question needed, and how much the whole model weighs. Raw bytes were counted on both sides — compressed is never compared against uncompressed.
See it on a real product.
Three models are published in full and open without an account. Ask one of them the question this page is about.