# OpenBMB's MiniCPM5-2B targets on-device AI

> The compact 2-billion-parameter model adds long-context handling and tool-calling in a footprint small enough to run locally.

Published by The Open Weights on Sep 6, 2026. Canonical: https://theopenweights.com/news/minicpm5-2b-30zi

## Key facts

- Company: OpenBMB
- Model: MiniCPM5-2B
- Version: 5
- Category: Text / LLM
- Modalities: Text / LLM
- License: Other (Open weights, commercial use allowed)
- Parameters: 2B
- Context window: 131K tokens
- Architecture: LlamaForCausalLM
- Precision: BF16
- Weights size: 5 GB
- Significance: notable
- Published: 2026-09-06
- Last verified: 2026-09-27
- Hugging Face: https://huggingface.co/openbmb/MiniCPM5-2B
- Canonical URL: https://theopenweights.com/news/minicpm5-2b-30zi

OpenBMB has released [MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B), the latest entry in its MiniCPM line of small language models designed to run efficiently on phones, laptops, and other resource-constrained hardware. At roughly 2 billion parameters, the dense text model is squarely aimed at developers who want capable inference without the cost or latency of cloud-scale systems.

According to the model card, MiniCPM5-2B pairs its compact size with two features that have become table stakes for practical deployments: long-context support and tool-calling. The former lets the model reason over larger documents and conversations, while the latter enables it to invoke external functions and APIs — a foundation for agentic workflows and structured output.

## Why it matters

The small-model space has grown crowded, but the appeal is consistent: a 2B model can be quantized and served on-device, keeping data local and eliminating per-token API fees. That makes MiniCPM5-2B a candidate for offline assistants, embedded applications, and privacy-sensitive settings.

- Roughly 2B parameters, dense (not a mixture-of-experts)
- Long-context and tool-calling support out of the box
- Distributed on Hugging Face under a custom license

Prospective users should review the licensing terms on the model card, which OpenBMB lists as a custom "other" license rather than a standard permissive option. As always with compact models, real-world performance will depend on the task, and teams should benchmark against their own workloads before committing.

## Get the model

- [Hugging Face](https://huggingface.co/openbmb/MiniCPM5-2B)

## Sources

- [openbmb/MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B) — Hugging Face, Sep 6, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "OpenBMB's MiniCPM5-2B targets on-device AI", The Open Weights, Sep 6, 2026, https://theopenweights.com/news/minicpm5-2b-30zi