CompressionStatus / live
Compressionlive
CaveGemma
Why use many token when few do trick — now baked in weights
A fine-tune of google/gemma-4-31B-it that speaks caveman without a system prompt or skill file. Across 193 held-out pairs it used 27% fewer output tokens, with code-fence exactness between 96.3% and 100%.
Product demo / CaveGemma
illustrative output comparisonprompt · no system prompt
Reverse a singly linked list in Python.
Flip each node next-pointer to prev. O(n) time, O(1) space.
```python
def reverse(head):
prev = None
while head:
head.next, prev, head = prev, head, head.next
return prev
```tokens
49144
saved
66%
illustrative estimate · ~4 chars/token
code-fence exactness96–100%
semantic similarity0.91–0.98
fewer output tokens27%
illustrative response pair · product figures from 193 held-out pairs
Capability ledger
06- 0127% fewer output tokens across 193 held-out pairs (weighted mean)
- 0296–100% code-fence exactness; semantic similarity 0.91–0.98
- 03Article density 0.005–0.020 on sources that measured 0.080
- 04Merged weights (62.5GB) or a 534MB LoRA adapter
- 05QLoRA rank 16; loads through Hugging Face transformers
- 06Text-only evaluation; multimodal behavior remains untested
Product ledger
02- Language
- Python
- License
- MIT · weights inherit Gemma terms
Product index
0501CavemanOutput compression for Claude Code & 30+ agents, plus the compression engine.Compression / live02Caveman CodeTerminal coding agent. Half the tokens.Agent toolkit / live03CavememPersistent memory your agents recall over MCP.Agent toolkit / live04CavekitCompressed, spec-driven development.Agent toolkit / live05Caveman ProxyThe byte-safe LLM gateway.Cloud / in development