A Tiny Open Model Built to Run on Your Phone

The Chinese research group OpenBMB has released MiniCPM5-2B, a language model small enough to run on a phone or laptop and open enough for anyone to use commercially. Published on 7 September under the permissive Apache 2.0 licence, the model has about 2.5 billion parameters, and its makers argue it now competes with systems several times larger.

According to a write up by MarkTechPost, MiniCPM5-2B averages 53.9 across 34 public benchmarks. The independent evaluation site Artificial Analysis places it at 15 on its Intelligence Index, one point behind a rival model that carries roughly three times the parameters. For something that fits comfortably on consumer hardware, that is a narrow gap.

Ordinary on the outside

Part of the appeal is how unremarkable the model is to deploy. It uses the standard Llama style architecture, so the common open source runtimes load it without custom code. The team lists support across llama.cpp, Ollama, LM Studio, Apple's MLX, vLLM and SGLang, which covers most of the ways developers actually run models locally today. It also carries a 128,000 token context window, once the preserve of far larger systems.

Alongside the main release, OpenBMB pushed a much smaller companion checkpoint of around 320 million parameters. Its only job is to help the larger model decode faster through a technique called speculative decoding, where a lightweight model drafts tokens that the main model then checks and accepts or rejects. The result is quicker responses on the same hardware.

Small is a strategy

The release fits a pattern that has defined much of 2026. While the headline race chases ever larger frontier models, a parallel effort keeps proving how much capability can be packed into something that runs offline, cheaply, and under a licence that lets companies ship it inside their own products. Open weights out of China have been central to that shift, and MiniCPM5-2B is another data point in the argument that most useful AI will not need a data centre to run.