Open-weights: intelligence vs size
Artificial Analysis Intelligence scores against model size, plus what runs on your hardware.
Intelligence vs parameters
What should I run?
Top picks for this hardware
All models in this view
Notes
- Sizes. Total parameters set memory, active parameters (MoE) set speed, and lookup parameters (n-gram or per-layer embedding tables) take memory but almost no bandwidth, so they can sit on SSD. The frontier ignores lookup tables, and frontier models are always shown.
- Fit. Weights fill VRAM, then RAM, then SSD if allowed. A model counts as runnable if a bandwidth roofline gives it a few tokens per second. Hardware figures are rough; see
assets/hardware.js. - Precision. 4-bit means NVFP4 on RTX 5000 / Blackwell and Q4_K_M elsewhere. AA scored hosted endpoints, so quantized local copies will score lower.
- Quadrant. Horizontal: coding and agentic percentile minus knowledge and instruction-following percentile (AA has no creative-writing eval). Vertical: speed, approximated by fewer active parameters. Dot size is the Intelligence Index, and a ring marks models that nothing beats on both speed and intelligence.
- Closed models. They publish no sizes, so they appear only as dashed score lines, one per product tier.
- Data. Scores, dates and parameter counts come from Artificial Analysis; exact sizes from Hugging Face. Lookup sizes and the Featured list are hand-curated in
data/.