MiniCPM5‑1B: Open‑Source 1B‑Parameter LLM for Edge AI and Tool‑Calling

MiniCPM5-1B, released by the OpenBMB team, is a dense 1.08 B‑parameter causal language model built on the standard LlamaForCausalLM architecture. It targets on‑device and resource‑constrained deployments while delivering 1B‑class open‑source state‑of‑the‑art performance, especially in tool‑use, code generation, and hard reasoning tasks. The model supports a 131,072‑token context window and features a hybrid reasoning mode toggled via a `<think>` chat template, allowing the same checkpoint to act as a fast assistant or a deliberative reasoner.

The model is distributed in multiple formats (BF16, GGUF for llama.cpp/Ollama, MLX for Apple Silicon, etc.) and integrates with popular inference backends such as vLLM, SGLang, and the Transformers library. SGLang is highlighted for native XML‑style tool‑calling, and the repository provides cookbooks for deployment and fine‑tuning across frameworks like TRL, LLaMA‑Factory, and ms‑swift. Training leveraged OpenBMB’s UltraData tiered datasets (Ultra‑FineWeb, Ultra‑FineWeb‑L3, UltraData‑Math, UltraData‑SFT‑2605) and a three‑stage post‑training pipeline (SFT → RL → On‑Policy Distillation) to boost reasoning accuracy while reducing over‑long responses.

MiniCPM5-1B’s long‑context capability, bilingual (English/Chinese) support, and edge‑AI friendliness make it a practical choice for local assistants, coding agents, and any application that needs a compact yet powerful LLM without relying on cloud APIs.