Architecture census — 271 models scanned · V1
Small Model Study
A pure-survey observation. We scanned all locally-available Ollama models on a 48-thread Xeon CPU with 157 GB RAM, kept the 88 with three billion parameters or fewer, and recorded their architecture metadata: family, parameter count, quantization, layer count, hidden dimension, attention heads, KV heads, FFN dimension, context length, RoPE frequency, GQA ratio, capabilities, thinking and tool support.
- scanned
- 271
- cataloged
- 88
- skipped (>3B)
- 183
- errors
- 0
Where the small-model mass sits
- under_100m
- 9
- 100m-500m
- 21
- 500m-1b
- 8
- 1b-2b
- 30
- 2b-3b
- 19
Model family distribution
- qwen2
- 16
- bert
- 14
- llama
- 14
- granite
- 7
- gemma3
- 4
- qwen3
- 3
- qwen35
- 3
- phi2
- 3
- stablelm
- 3
- nomic-bert
- 2
- glmocr
- 2
- granitemoe
- 2
- gemma
- 2
- gemma2
- 2
- exaone
- 2
- minimax-m2
- 1
- gemini
- 1
- nomic-bert-moe
- 1
- lfm2
- 1
Quantization distribution
- Q4_0
- 27
- F16
- 22
- Q4_K_M
- 20
- Q8_0
- 11
- BF16
- 3
- fp16
- 1
- FP8
- 1