MiniCPM

OpenBMB is the open-source arm around Tsinghua's NLP group and ModelBest (面壁智能), and MiniCPM is its on-device line, nicknamed 小钢炮 for a reason. The current small-model flagship is MiniCPM5-1B, released 25 May 2026: 1,080,632,832 parameters total with 679M non-embedding, 24 layers, 16 query heads against 2 key-value heads, a 131,072 token context and Apache 2.0 licensing on weights and code alike.

The design choice worth studying is the single-checkpoint dual mode. An enable_thinking flag switches the same weights between a fast assistant and a deliberate reasoner emitting think traces, so an edge device ships one model instead of two and saves the storage that matters most on constrained hardware. OpenBMB reports a 42.57 average across its thinking benchmark set against 26.77 for Qwen3-0.6B and 35.61 for LFM2.5-1.2B, crediting reinforcement learning plus its OPD stage, alongside a 29 point drop in responses that hit the max token limit. Independent measurement is more restrained: Artificial Analysis places MiniCPM5-1B at an intelligence index of 11.7, so calibrate expectations to a strong 1B rather than a small frontier model.

Deployment is where the family earns its place in a catalogue. GGUF for llama.cpp and Ollama, MLX and 4-bit builds for Apple Silicon, plus FlagOS adaptations covering Nvidia, Ascend, Hygon, Metax and Iluvatar. That last item is what English-language coverage tends to skip and it is the deciding factor if your device is going into a mainland product where Nvidia is not the procurement default. The siblings are worth knowing too. MiniCPM-V 4.6, released 11 May 2026, is a 1.3B multimodal model built on SigLIP2-400M and Qwen3.5-0.8B that handles single-image, multi-image and video understanding on iOS, Android and HarmonyOS with vLLM, SGLang, llama.cpp and Ollama support at launch. MiniCPM-SALA, from 11 February 2026, is a sparse and linear hybrid attention model aimed at million-token context. Weights for all of them land on Hugging Face and ModelScope simultaneously.