OpenBMB/MiniCPM

What it solves

MiniCPM is a series of compact, high-performance language models designed for on-device, local deployment and resource-constrained environments. It aims to provide state-of-the-art (SOTA) performance in the 1B to 2B parameter class, making powerful AI capabilities like coding, mathematics, and agentic tasks accessible on edge devices without requiring massive cloud infrastructure.

How it works

The project utilizes a dense Transformer architecture (specifically LlamaForCausalLM) and a sophisticated training pipeline called UltraData Tiered Data Management. This process involves three stages:

1. Base Training: Establishes core language stability and capability.
2. Mid-training: Strengthens specific target capabilities and adapts the model to target data distributions.
3. Post-training: Employs Supervised Fine-Tuning (SFT) with deep-thinking data, Reinforcement Learning (RL) using critic-based algorithms, and On-Policy Distillation (OPD) to merge capabilities from multiple expert RL models into a single release model.

Who it’s for

It is intended for developers building local assistants, coding agents, and tool-use workflows, as well as users who need high-reasoning capabilities in environments with limited computational resources.

Highlights

• 2B-class SOTA: MiniCPM5-2B outperforms other open-source models of similar size and remains competitive with 4B-class models.
• On-Device Optimization: Specifically built for local deployment and resource-constrained scenarios.
• Long-Context Support: Native support for context lengths up to 131,072 tokens.
• Open High-Quality Data: Releases accompanying training datasets (UltraX, UltraData-Code, etc.) to support the community.
• Broad Deployment Support: Compatible with vLLM, SGLang, Llama.cpp, and Transformers.