MiniCPM5-2B: New 2B Model for Local Deployment Released on HuggingFace

On September 6, 2026, the model openbmb/MiniCPM5-2B with the text-generation task was published on HuggingFace. It is the second model in the MiniCPM5 series after MiniCPM5-1B — a dense 2B transformer designed for local deployment and resource-constrained scenarios. The authors claim SOTA performance among open 2B-class models and competitiveness with 4B-class models in coding, math, long-context understanding, and agentic tasks.

What it means

The model weighs 4.7 GB in safetensors format, contains 2.5 billion parameters, and is distributed under the apache-2.0 license. It runs via transformers. The platform already hosts 16 quantized builds in GGUF and MLX formats; the most compact working one is IQ4_XS at 1.3 GB by PollardWeights. For local running on CPU or GPU with limited VRAM, this lets you pick a build that matches your available hardware: from the full model down to a 4-bit quant. This release is of interest to anyone looking for a compact text model for local use: the model weights are open, the license is permissive, and quantized variants are already available from the community. You can keep track of similar model releases in the CPU3D news feed.