MiniCPM Text Series
The MiniCPM text series is a family of high-performance language models built for on-device deployment, with parameter sizes ranging from 0.5B to 9B.
Series overview
The MiniCPM text series keeps evolving around efficient on-device inference, long context and agent capabilities. It currently includes MiniCPM5, MiniCPM-SALA and MiniCPM4 / 4.1.
MiniCPM5 A new generation of compact agent models that combine long context, Think / No Think hybrid reasoning and tool calling.
MiniCPM-SALA Built for million-token ultra-long context. It uses hybrid sparse and linear attention, with a focus on optimizing long-sequence inference efficiency.
MiniCPM4 / 4.1 Covers on-device deployment at 0.5B and 8B, with sparse attention, hybrid reasoning and on-device generation acceleration.
| Model | Parameters | Context | Highlights |
|---|---|---|---|
| MiniCPM5-2B | 2B | 128K | Strong performance on code, math, long text, tool calling and agent tasks |
| MiniCPM5-1B | 1B | 128K | Hybrid Think/No Think reasoning, tool calling |
| MiniCPM-SALA | 9B | 1M | Hybrid sparse + linear attention |
| MiniCPM4.1-8B | 8B | 128K | Sparse attention, hybrid reasoning |
| MiniCPM4-8B | 8B | 128K | 5x on-device generation speedup |
| MiniCPM4-0.5B | 0.5B | 128K | Ultra-lightweight on-device model |
Recommended choices
| Use case | Recommended model | Why |
|---|---|---|
| Local assistants, coding agents, tool calling | MiniCPM5-2B, MiniCPM5-1B | Compact 1B scale; the same weights support Think / No Think, with native 128K context and tool-calling capability. |
| Million-token documents, ultra-long text analysis | MiniCPM-SALA | Supports up to 1M context; hybrid sparse and linear attention is well suited to ultra-long-sequence inference. |
| 8B-class deep reasoning and long-text tasks | MiniCPM4.1-8B | Supports sparse attention and hybrid thinking, letting you switch flexibly between reasoning quality and runtime efficiency. |
| Mature general-purpose on-device deployment | MiniCPM4-8B | 128K context with on-device generation acceleration, and support for many mainstream inference frameworks. |
| Severely resource-constrained, ultra-lightweight applications | MiniCPM4-0.5B | The smallest model in the series, suited to scenarios where storage, memory and compute costs come first. |
Quick advice: If you're unsure, start with MiniCPM5-2B and MiniCPM5-1B; choose MiniCPM-SALA if you need million-token context; choose MiniCPM4.1-8B if you need stronger 8B-class capability.
See the sub-pages for detailed deployment and fine-tuning guides.
Related Resources
| Topic | Link |
|---|---|
| Code repository and release notes | OpenBMB/MiniCPM |
| MiniCPM5-2B model weights | Hugging Face / ModelScope |
| MiniCPM5-1B model weights | Hugging Face / ModelScope |
| MiniCPM-SALA model weights | Hugging Face / ModelScope |
| Model documentation | MiniCPM5-2B / MiniCPM5-1B / MiniCPM-SALA / MiniCPM4.1 / MiniCPM4 |
| Agent Skills | Deployment and fine-tuning skills |
| Online demo | MiniCPM5-2B Demo |
| Technical reports | MiniCPM4 / MiniCPM-SALA |
| Text model technical community group | Join the MiniCPM text model community |