Compute Costs

Compute Costs

Breakdown

Layer	Cost	What
Outer loop (ultrawhale v1→v100)	$37.19	DeepSeek API — the coding agent that built itself, generated the dataset
Inner loop (kompress v3→v17)	$1.76	vast.ai RTX 4090 — 17 models, 8 teachers, 4 architectures
Total	$38.95	Less than a single conference registration

Per-model cost

• v3 training: ~$0.09 (15 min RTX 4090)
• v4 self-label + train: ~$0.15 (25 min)
• v5 self-label + train: ~$0.15 (25 min)
• v8 C3 distillation: ~$0.20
• Average per model: ~$0.10

The economics

• $0.24 per release across 150 blocks (ultrawhale)
• $0.10 per model across 17 models (kompress)
• The entire research program cost less than a single industry conference registration
• No institutional grants, no corporate sponsorship — personal expenditure + decentralized open-source funding