MiniCPM5-1B
by OpenBMB
A 1 billion parameter LLM that punches well above its weight class for local and on-device AI deployment.
See https://huggingface.co/openbmb/MiniCPM5-1B
Features
- 1 billion parameter dense LLM, Llama-style causal architecture
- Compatible with Ollama, LM Studio, vLLM, and MLX
- Hybrid reasoning mode: enable a “thinking switch” for harder prompts that require step-by-step reasoning
- Just over 2 GB VRAM usage — runs on consumer GPUs and even CPU (slower)
- Gradio interface for browser-based access on port 7860
- Training recipe disclosed: pre-training, supervised fine-tuning, RL + OPT (on-policy distillation)
- Coherent responses and reasonable performance on constrained coding tasks
- HTML canvas coding demo capability (single-file apps)
Superpowers
MiniCPM5-1B is exceptional for its size. Most 1B models are too limited for practical use, but MiniCPM5 demonstrates unusually strong reasoning and coding ability at this scale, driven by OpenBMB’s transparent training methodology combining pre-training, SFT, and reinforcement learning with on-policy distillation. At just over 2 GB VRAM, it runs on nearly any modern consumer GPU or even CPU, making it the go-to choice for truly local, always-on AI on modest hardware. The hybrid reasoning mode is a standout — it lets the model pause and think on harder prompts, something almost no 1B model supports. Weaknesses: multilingual performance is poor and results outside English are unreliable.
Pricing
Open weights — free to use.