OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device

AITopTools Editorial TeamSeptember 7, 2026

What changed

OpenBMB has released MiniCPM5-2B, a 2.52-billion-unit AI model designed to run on devices and handle long inputs. The release includes model files, training data, and earlier development versions under the Apache 2.0 open license, with builds starting at 1.56 GB and support for several popular AI tools.

What this means for you

Developers can download and run the model locally through tools including Ollama, llama.cpp, vLLM, SGLang, and MLX, without changing the model code. Its relatively small file size may make local use more practical, though the feed does not state which consumer devices can run it or how fast it will be.

Related AI news

Read Google Releases Gemini 3.8 Live and 3.8 Live Extended Thinking for Production Grade Voice Agents
New AI featuresSep 15, 2026

Google Releases Gemini 3.8 Live and 3.8 Live Extended Thinking for Production Grade Voice Agents

The models are available now in Google’s Gemini API and AI Studio, allowing developers to build voice applications at $0.005 per minute for audio input. Generated audio includes Google DeepMind’s SynthID watermark, which identifies it as AI-created.

MarkTechPostSee why it matters
Read Meta now lets AI agents handle the boring parts of WhatsApp Business setup
AI toolsSep 15, 2026

Meta now lets AI agents handle the boring parts of WhatsApp Business setup

Developers can use tools such as Claude, Cursor, Codex, and ChatGPT to reduce the manual work involved in launching WhatsApp Business messaging. The feature is aimed at developers, and the feed does not specify pricing or broader access details.

TechCrunch AISee why it matters