We build Ralph, Bittensor SN40's open model-compression tournament. Miners compress a pinned model without changing its architecture; fresh public rounds select a winner in each bit tier; the current crowns ship as exact GGUF files anyone can download and check.
- Run a crown: https://huggingface.co/RalphLabsAI/ralph-crowns
- Inspect Round 7: https://huggingface.co/datasets/RalphLabsAI/ralph-v2-rounds/blob/main/rounds/round-00000007-f42ae20c6ed2a796.json
- Read or build Ralph: https://github.com/RalphLabsAI/ralph
- Visit the project: https://ralphlabs.ai
Four current tiers: binary · ternary · sub2 · sub4.
Ralph retention is a protocol-specific compression-fidelity measure, not a general capability benchmark. Device claims are published only with a reproducible receipt.
Looking for the former training-recipe system?
ralph-v1andrecipe-v1preserve that archived history.