OpenBMB put its full MiniCPM5-1B stack in the open Monday: weights, the pre-training corpus, and deployment code, all under Apache-2.0. The 1B-parameter model landed on Hugging Face with base, SFT, GGUF, and Apple Silicon checkpoints.
What separates it from the usual "open weights" drop is the data. The training corpus ships alongside the model as Ultra-FineWeb, plus math and SFT sets, so the recipe is meant to be reproducible rather than just runnable. Cookbooks for vLLM, SGLang, llama.cpp, and Ollama sit in the GitHub repo.
OpenBMB claims the top spot on the Artificial Analysis small-model index, scoring 17.9 to edge the larger Qwen3.5-2B at 16.3. That number is company-reported and hasn't been independently confirmed. The model card itself frames the lead more narrowly, as best within its own comparison set of sub-2B models.
The practical pitch is size. The 4-bit build runs around 0.5GB on phones, laptops, and in browsers, with one checkpoint that flips between fast chat and a slower thinking mode. It also drives a local "desktop pet" that needs no cloud. Tool-calling support is wired into SGLang now.
Bottom Line
MiniCPM5-1B is a dense 1.08B model whose 4-bit build is about 0.5GB, with the training corpus released alongside the weights.
Quick Facts
- Model: MiniCPM5-1B, dense Transformer, 1,080,632,832 parameters
- License: Apache-2.0
- Context length: 131,072 tokens
- 4-bit build: roughly 0.5GB, runs on phones and in browsers
- Artificial Analysis score: 17.9 vs Qwen3.5-2B 16.3 (company-reported)
