GitHub radar
MiniCPM5-2B — a local model to handle daily tasks
OpenBMB just launched MiniCPM5-2B, an on-device Ollama/LM studio friendly model. Here’s how good it really is.
MiniCPM5-2B is a 2B parameter model by OpenBMB (part of their MiniCPM5 series), meant for local and on-device use. According to OpenBMB, it’s the “best open model of its size” in their tests and even rivals 4B models. It’s good at coding, math, long context, tools, and agency. The repo for the GGUF files for llama.cpp, Ollama and LM studio can be found here. The compressed Q4_K_M version is ~1.6GB and the 8-bit version is ~2.7GB. OpenBMB also released the training data for this model and the weights are available under the Apache 2.0 license.
Why a vibe-coder should care
So how good is this model? Well, it’s a 2B parameter model, so it’s nowhere near my daily driver Opus 5 in Claude Code, but according to OpenBMB, it’s the best open model of its size in their tests. Still, it’s only 2B parameters. I’d only recommend using this model for local experimentation and better understanding of local models, or for simple and repetitive tasks such as sorting text or extracting a few attributes. Definitely not for any kind of real agency, and don’t expect to get all those “free agents on your device” things. Bonus points for being Apache 2.0 licensed and having open data.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up MiniCPM5-2B locally from https://huggingface.co/openbmb/MiniCPM5-2B-GGUF — use the Q4_K_M version, run it with Ollama and let me try one simple task on it.
A regular laptop with 8 GB of RAM or more — the compressed version takes around 1.6 GB.
Open on Hugging Face▌ More finds