on-device
taxonomy axis: design_intent
Decision rule: Framed for edge/deployment efficiency constraints.
Anchor: Phi-3-mini
Models in this category (4)
| Model | Org | Date | Block | Attention | Context | Total | Active |
|---|---|---|---|---|---|---|---|
| Phi-3 | Microsoft | 2024-04-22 | dense | MHA | 4,096 | 3.8B | 3.8B |
| Gemma 3 | Google DeepMind (Gemma Team) | 2025-03-25 | dense | sliding-window | 131,072 | 27B | 27B |
| Tiny Aya | Cohere Labs / Cohere | 2026-03-12 | dense | hybrid | 8,192 | 3.35B | 3.35B |
| Gemma 4 | Google DeepMind (Gemma Team) | 2026-07-24 | dense | hybrid | not disclosed | 31.25B | 31.25B |
What varies within this category
attention variant varies — MHA, hybrid, sliding-window. trained context varies — 131072, 4096, 8192, not disclosed. openness varies — open-weights, undisclosed. scale class varies — large, medium.