OpenBMB ships MiniCPM5-2B for local agents

OpenBMB released MiniCPM5-2B, a 2.52-billion-parameter dense model with a 131,072-token context window for on-device assistants, coding agents, and tool-use workflows, according to its official model card. The Apache-2.0 release includes BF16, GGUF, MLX, GPTQ, and LiteRT variants and uses standard LlamaForCausalLM, lowering runtime-specific integration work without making benchmark claims independent. Teams comparing small-model deployment floors should treat it as a local test candidate, then measure accepted-task quality, memory use, and latency on their target device before displacing a larger model.