Ralph

Organizations deploying large language models face a critical tradeoff: serving Qwen/Qwen3-8B at full precision consumes substantial bandwidth and compute resources, while naive quantization degrades downstream capability unpredictably. Existing compression approaches either require architectural changes that break compatibility or use imitation-based scoring that rewards surface-level style matching rather than retained capability—leading to models that score well on narrow metrics while failing on real tasks.

A decentralized network of compression researchers can explore the full solution space without coordination overhead, each optimizing for explicit bit budgets in their method of choice. By measuring retention through downstream observer effects rather than teacher imitation, a tournament structure rewards genuine capability preservation and makes copying economically irrational: an exact duplicate has zero margin over what it copied and cannot earn challenger share. The commit-then-generate mechanism prevents exam harvesting, while publicly signed, recomputable round records let any participant verify the scoring without trusting the operator.

Ralph implements this on Bittensor SN40 using three dimensions: four simultaneous bit-budget tiers (binary at 1.15 bits/weight, ternary at 1.75, sub2 at 2.3, sub4 at 4.0) where miners can enter once per round, strict intake gates that measure actual serialized bits rather than declared types, and scored competition where the displayed metric reflects effect magnitude and effect similarity to the pinned parent across three judge models. Crown status—80% of tier allocation—holds for three rounds or changes on a displayed-metric margin of 0.02 in one round or 0.01 in two consecutive rounds, with a requirement that no challenger slice falls below the incumbent's worst slice.

Ralph SN40 v2 is an architecture-preserving model compression subnet on Bittensor where miners compress the pinned Qwen/Qwen3-8B model into four bit-budget tiers (binary, ternary, sub2, sub4) and compete for crown status. Scoring measures retention by downstream effect using three judge models rather than imitation, with publicly signed, re-runnable round records and strict bit-budget enforcement via GGUF tensor types or safetensors quantization levels.