OpenBMB Ships MiniCPM-4 — The 8B Model That Claims to Match Models 10x Its Size

OpenBMB — the open-source AI lab behind the MiniCPM model family — shipped MiniCPM-4 this week, the fourth generation of their efficient small model series. The release is significant: the 8B flagship model benchmarks comparably to 30B+ models on standard reasoning evals, while the 0.5B variant runs on hardware as constrained as a mid-range Android phone.

The architectural novelty in MiniCPM-4 is a hybrid inference system that dynamically switches between dense and sparse activation depending on the complexity of the current token prediction. For simple continuations (common words, predictable patterns), the model activates fewer parameters. For harder reasoning steps, it activates more. This is similar in spirit to mixture-of-experts, but applied at inference time rather than through a learned router — the model decides dynamically rather than being assigned to experts at training.

MiniCPM-4 is released under Apache 2.0 with full weights for the 0.5B, 2B, 4B, and 8B variants. Quantized versions run on Android via the team's MLC-LLM integration, with reported 40-token/second throughput on a Snapdragon 8 Elite device. The same team also released VoxCPM2 alongside this drop — a tokenizer-free TTS system that uses MiniCPM-4 as its language backbone, effectively shipping an end-to-end voice assistant stack in a single open-source release.

The Chinese open-source AI lab ecosystem continues to produce models that punch significantly above their parameter count. MiniCPM-4 follows the trajectory set by Qwen 3 and DeepSeek R2 — efficient architectures closing the gap with frontier models at a fraction of the compute cost. For developers building on-device AI features for mobile apps, the MiniCPM-4 stack (language + voice) is now a credible full-featured option without any API calls.