Beijing – September 12, 2026 -- ModelBest, a Chinese AI startup, has released MiniCPM5-2B, a 2-billion-parameter open-source language model that ranks first on the Artificial Analysis Intelligence Index among open-source models under 4 billion parameters. The model, built in collaboration with the OpenBMB open-source community, also scored 20 on the Agentic Index, a metric measuring autonomous task-execution capability, outpacing peers in its size category.
Model runs agentic tasks natively on edge hardware without cloud dependency
MiniCPM5-2B natively supports tool calling, deep search, code generation, and multi-step reasoning despite its compact parameter count. ModelBest positions the release as evidence that increasing 'intelligence density' rather than scaling parameter counts can deliver strong performance within the compute, memory, and power constraints typical of edge devices such as PCs, smartphones, robotics, and IoT hardware.
Full technical stack released, not just model weights
Alongside the model weights, ModelBest and OpenBMB published the complete training pipeline, including datasets, training recipes, and reinforcement learning infrastructure. This end-to-end transparency covers data curation, pre-training, and alignment stages, giving developers a reproducible framework to fine-tune and deploy specialized edge applications rather than working with opaque, weights-only releases.
Local deployment cuts latency, API costs, and privacy exposure
Running document processing, data synthesis, code generation, and multi-turn Q&A directly on-device eliminates dependency on cloud APIs, according to ModelBest. The company states this shift reduces latency and infrastructure costs while keeping data processing local, addressing privacy and data sovereignty concerns for enterprise users.
MiniCPM family has logged over 50 million downloads to date
ModelBest reports cumulative downloads across its MiniCPM model family have surpassed 50 million. The company frames the release of models, methodologies, and training recipes as part of a broader push to expand access to lightweight, deployable edge AI globally.