MiniCPM4 series (8B flagship)
On-device LLM family claiming over 5x generation speed-up on typical edge chips; the 8B flagship was trained on 8T tokens.
Designed around four efficiency levers (architecture, data, training algorithms, inference system). Apache-2.0.
- Date
- Friday, 6 June 2025
- Lab
- ModelBest (OpenBMB)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Generation speed-up on end-side chips | over 5x company-claimed | company |
| MiniCPM4-8B training tokens | 8T | company |
Date from the model card (2025.06.06).
Sources
This record was checked against its sources on 6 October 2026. How we check