scratchy Models Docs
llama mistral phi3 qwen2 qwen3 mixtral qwen3-moe granite qwen2-moe commandr gemma2 qwen3-5 deepseek-v2 deepseek-v3-flat qwen3-5-moe gemma3 deepseek-v3 modernbert gemma4 gemma3-mm gemma4-moe qwen2-5-vl qwen2-vl qwen3-5-vl locateanything

Scratchy model architectures

25 architectures, 1,025 lines of DSL between them. Each one is the model's math, written once; the compiler turns it into the code for every target. Pick a model from the list on the left, then optionally diff it against another below.

— none — llama — 22 lines mistral — 22 lines phi3 — 22 lines qwen2 — 25 lines qwen3 — 24 lines mixtral — 20 lines qwen3-moe — 22 lines granite — 23 lines qwen2-moe — 23 lines commandr — 36 lines gemma2 — 56 lines qwen3-5 — 46 lines deepseek-v2 — 47 lines deepseek-v3-flat — 46 lines qwen3-5-moe — 53 lines gemma3 — 34 lines deepseek-v3 — 48 lines modernbert — 52 lines gemma4 — 48 lines gemma3-mm — 74 lines gemma4-moe — 57 lines qwen2-5-vl — 58 lines qwen2-vl — 54 lines qwen3-5-vl — 53 lines locateanything — 60 lines