MiniCPM-4 8B

MiniCPM4-8B is the flagship of OpenBMB's MiniCPM4 series, released June 6, 2025, and designed explicitly for end-side devices. It packs 8 billion parameters trained on 8 trillion tokens, with systematic efficiency innovations across architecture, training data, training algorithms, and inference systems.

The headline innovation is InfLLM-v2, a trainable sparse attention mechanism where each token computes relevance with less than 5 percent of tokens in 128K long-context processing, dramatically cutting the cost of long-text inference. Combined with hybrid reasoning and frequency-ranked speculative decoding, the series claims over 5x generation acceleration on typical end-side chips while maintaining best-in-class performance at its scale.

The model is open-weight under the Apache 2.0 license, runs on Hugging Face transformers, vLLM, SGLang, llama.cpp, Ollama, and CPM.cu, and sits in a family that also includes a 0.5B version, a survey-paper agent variant, and an MCP-tool-calling variant.