MiniCPM5-2B is a compact text model from OpenBMB for local assistants, coding agents, tool use, and reasoning on constrained hardware. Despite the short name, the checkpoint contains about 2.52 billion total parameters, with about 1.98 billion non-embedding parameters.
It is not a multimodal MiniCPM-V model. MiniCPM5-2B accepts text and produces text; choose a MiniCPM-V or MiniCPM-o checkpoint if image, video, or audio input is essential.
Specifications
| Item | MiniCPM5-2B |
|---|---|
| Architecture | Dense LlamaForCausalLM |
| Parameters | 2,516,756,480 total; 1,981,982,720 non-embedding |
| Layers | 42 |
| Context | 131,072 tokens |
| Modalities | Text input and output |
| Languages highlighted | English and Chinese |
| Licence | Apache 2.0 |
| Formats | BF16, GGUF, MLX, GPTQ, SFT, Base and other project releases |
OpenBMB calls it a 2B-class model because of the non-embedding count. Hugging Face may display it as 3B after rounding the total parameter count. Both labels refer to the same checkpoint.
What it is designed to do
The official release emphasizes local assistants, coding agents, tool-use workflows, and reasoning. A 131K context window makes it unusually flexible for its size, but context length is a capacity ceiling—not proof that a small model will reason reliably over an entire repository.
MiniCPM5-2B is most interesting for:
- local command or retrieval assistants with narrow responsibilities;
- inexpensive tool-selection and structured tasks;
- agent experiments where privacy or offline use matters;
- compact coding assistance with validation around the output;
- edge deployments where a 7B–14B model is too large.
It is less suitable as an unsupervised replacement for a frontier coding model. Small models need tight prompts, limited tool permissions, schema validation, and deterministic checks.
Running the GGUF build
OpenBMB publishes an official GGUF repository for llama.cpp, Ollama, and LM Studio. With current llama.cpp tooling:
llama serve -hf openbmb/MiniCPM5-2B-GGUF:Q4_K_MThat starts an OpenAI-compatible local server. A Q4 build materially reduces memory compared with BF16, but exact RAM and VRAM use depends on the quant file, context allocation, cache type, and offload settings. Avoid quoting one universal hardware requirement.
For Apple Silicon, OpenBMB also publishes an MLX/4-bit build. NVIDIA deployments can use Transformers, SGLang, or the project’s other supported formats. See our local AI hub, VRAM guide, and Ollama guide for the surrounding setup decisions.
Tool use and agent safety
OpenBMB reports tool-use and coding results in its own evaluation tables. Treat those as project benchmarks rather than independently reproduced guarantees. For a production agent:
- expose a small allowlist of tools;
- validate arguments before execution;
- require approval for destructive actions;
- test failure and retry paths;
- compare the exact quantization you plan to deploy.
Quantization can change instruction following, coding accuracy, and tool syntax. A result from BF16 does not automatically describe Q4 performance. Use the AI Testing & Evaluation hub to build a representative evaluation set.
MiniCPM5-2B versus larger local models
Choose MiniCPM5-2B when footprint, local privacy, or inexpensive concurrency dominates. Choose a larger model when multi-file coding, ambiguous instructions, long reasoning chains, or autonomous operation matter more than resource use.
PAIR can distribute independent requests across prepared machines, but it cannot combine their memory to fit a larger model. Read the NVIDIA PAIR guide if you want several local nodes to serve concurrent jobs.