This 2-billion-parameter AI model runs on your phone — and just beat every sub-4B competitor

Facebile Intelligence — the Beijing-based AI lab behind the MiniCPM family of on-device models — has a new entry that punches well above its weight class. At the 2026 World AI Conference on Sunday, the company teamed up with OpenBMB to launch MiniCPM5-2B, a 2-billion-parameter model that tops the Artificial Analysis (AA) Index with a score of 17, the highest ever recorded for any model under 4 billion parameters.

What makes the number worth noting is what it measures: the AA Index is an independent benchmark that scores models across general knowledge, math, code reasoning, instruction following, multilingual understanding, long-text processing, tool calling, and deep search. For a model designed to run locally on phones, laptops, and cars rather than in a data center, covering that range at this size is unusual.

MiniCPM5-2B comes with native hybrid thinking — it can toggle between fast, low-latency responses and deeper, chain-of-thought reasoning depending on the task. The context window stretches to 512K tokens, enough to ingest a full-length novel or hundreds of thousands of lines of code in a single pass.

The model was built from the ground up as an on-device agent. The training pipeline ran through 200 billion tokens of agent-focused mid-training, followed by supervised fine-tuning on a million high-quality agent trajectories, and capped with scalable reinforcement learning alignment. The result is a model that can call tools, search the web, write and execute code, and act as the reasoning engine inside a local AI assistant without phoning home to a server.

Facebile Intelligence has already lined up broad hardware support. On the global side, MiniCPM5-2B has been adapted for chips from AMD, Intel, MediaTek, and Qualcomm. Domestically, the company worked with the FlagOS community to secure Day-0 support across nine Chinese chip platforms including Huawei Ascend, Hygon, Kunlunxin, Mthreads, Moore Threads, T-Head, TsingMicro, Enflame, and NVIDIA, with ARM-side platforms also covered.

The company says MiniCPM5-2B will be open-sourced in the near future. Weights will be available through Hugging Face and ModelScope, along with deployment images and guides for each supported chip platform. Given the MiniCPM5-1B predecessor already ran directly in a browser, the 2B variant should find its way onto a wide range of consumer devices quickly.