Caveman

03

Labs

Our own models, and the research behind them. Everything here is measured before it is claimed.

Model cardopen weights

CaveGemma

fine-tune of google/gemma-4-31B-it · QLoRA r16 · MIT · weights inherit Gemma terms

Caveman compression baked into Gemma's weights.

Get the weights →
27%
fewer output tokens · 193 pairs
96–100%
code-fence exactness
0.91–0.98
semantic cosine
534MB
LoRA adapter
Research4 papers
  1. 01Logs as pixelsText into PNG. When it wins, when it loses.Read →
  2. 02The caveman phenomenon73k stars for a prompt that drops articles.Read →
  3. 03CaveBench methodologySavings at held quality. Cost per correct task.Read →
  4. 04The zero-dollar dashboardWhy verified savings starts at nothing.Read →