MiniCPM Models — Hardware Requirements
12 MiniCPM models from OpenBMB and the community, from the smallest that runs in 0.5 GB of VRAM up to 8x2B parameters. Every row links to full quantization tables and GPU compatibility.
All MiniCPM Models by Size
| Model | Params | Runs from | Context | Publisher | Quant downloads |
|---|---|---|---|---|---|
| MiniCPM5 2B DSpark | 324M | 0.5 GB | 131K | ||
| MiniCPM5 1B CoreAI | 1B | 2.2 GB | — | ||
| MiniCPM5 1B Claude Opus Fable5 v2 Thinking | 1.1B | 0.8 GB | 131K | ||
| MiniCPM5 1B Agentic Tooluse Merged FP16 | 1.1B | 0.8 GB | 131K | ||
| Huihui MiniCPM5 1B Abliterated | 1.1B | 0.6 GB | 131K | ||
| MiniCPM5 1B | 1.1B | 0.6 GB | 131K | ||
| MiniCPM5 1B Base | 1.1B | 0.8 GB | 131K | ||
| MiniCPM5 1B SFT | 1.1B | 0.8 GB | 131K | ||
| MiniCPM5 1B Claude Opus Fable5 v2 Thinking Heretic | 1.1B | 0.8 GB | 131K | ||
| MiniCPM V 4.6 | 1.3B | 0.9 GB | 262K | ||
| MiniCPM V 4.6 Thinking | 1.3B | 0.7 GB | 262K | ||
| MiniCPM 2B Sft BF16 | 2B | 1.9 GB | 4K | ||
| MiniCPM5 2B | 2.5B | 1.2 GB | 131K | ||
| MiniCPM5 2B Base | 2.5B | 1.5 GB | 524K | ||
| MiniCPM5 2B SFT | 2.5B | 1.5 GB | 131K | ||
| MiniCPM5 2B Midtrain | 2.5B | 1.5 GB | 131K | ||
| Minicpm5 2B Distilled Reasoning | 2.5B | 1.5 GB | 131K | ||
| MiniCPM O 2 6 | 8.7B | 3.4 GB | 33K | ||
| MiniCPM V 4 5 | 8.7B | 4.3 GB | 41K | ||
| MiniCPM O 4 5 | 9.4B | 4.6 GB | 41K | ||
| MiniCPM SALA | 9.5B | 19.3 GB | 524K | ||
| MiniCPM MoE 8x2B | 8x2B | 7.8 GB | 4K |
Frequently Asked Questions
- How much VRAM do I need to run a MiniCPM model?
- The smallest MiniCPM model, MiniCPM5 2B DSpark, runs from 0.5 GB of VRAM at an aggressive quantization. Larger family members need proportionally more — see the table above for every model.
- Which MiniCPM models can I run on a 16 GB GPU?
- 21 of 22 MiniCPM models fit in 16 GB of VRAM at some quantization, including MiniCPM5 1B Claude Opus Fable5 v2 Thinking, MiniCPM O 4 5, MiniCPM V 4 5.
- What is the most popular MiniCPM model to run locally?
- MiniCPM SALA is the most downloaded MiniCPM model in local-friendly quantized formats. It runs from 19.3 GB of VRAM.