Categories

One value per taxonomy axis. Every model gets exactly one value per axis; a model that fits no value forces the taxonomy to evolve, not the model to fit.

compute structure

• dense (16)
  Every transformer layer's FFN is fully active per token; no routing anywhere.
• sparse-MoE (27)
  MoE layers dominate (≥80% of transformer layers are routed FFNs).
• hybrid (3)
  Neither pattern reaches 80% of layers (e.g. a substantial mix of dense and MoE layers).

sequence mixing

• full-attention (18)
  All layers use unrestricted attention over the whole context.
• efficient-attention (9)
  Attention with restricted/computed patterns (sliding window, linear attention, sparse patterns) — but still attention.
• SSM (2)
  State-space layers (Mamba-style) do the mixing; no attention in the block.
• hybrid (14)
  Mixed attention + SSM/linear layers (or attention variants in a deliberate per-layer pattern).

modality

• text-only (41)
  Report describes no non-text input or output.
• multimodal (5)
  Report describes vision/audio/other modalities in or out.

openness

• open-weights-open-data (6)
  Weights **and** training data released.
• open-weights (34)
  Weights released; data not released.
• open-data (0)
  Data released; weights not.
• closed (0)
  Report says the model is not released.
• undisclosed (6)
  Report is silent on release. Absence is itself the classification.

scale class

• frontier (22)
  ≥ 100B
• large (11)
  10B ≤ p < 100B
• medium (12)
  1B ≤ p < 10B
• small (0)
  < 1B

design intent

• frontier-generalist (25)
  Framed as a general-purpose foundation model; broad benchmark coverage.
• reasoning-specialised (10)
  Framed around reasoning/thinking (CoT, RLVR, test-time compute).
• on-device (4)
  Framed for edge/deployment efficiency constraints.
• long-context (3)
  Framed primarily around context length.
• domain-specific (4)
  Framed for a specific domain (code, biology, law…).