OpenBMB released MiniCPM5-2B under Apache 2.0 for on-device use
A 2-billion-parameter model scoring 53.9 on average across benchmarks, aimed at local assistants, coding and tool use. The licence is the part that matters.
OpenBMB has released MiniCPM5-2B, a two-billion-parameter model licensed under Apache 2.0 and aimed at on-device use — local assistants, coding and tool calling. It scores 53.9 on average across the benchmark suite reported at release.
A small model with a permissive licence is not a frontier event, and the benchmark average will be beaten. The licence is the durable part: Apache 2.0 permits commercial use, modification and redistribution without the field-of-use restrictions that most "open" model licences carry.
Why 2B is a deliberate size
Two billion parameters is chosen to fit constraints rather than to compete on capability. Quantised, a model this size runs on a phone, a laptop without a discrete GPU, or an embedded device, within a memory budget that leaves room for the application using it.
That makes it a different product from a frontier model rather than a worse one. The tasks it is aimed at — classification, extraction, routing, structured tool calls, local assistance — are the high-volume, low-complexity work that makes up most of what production systems actually ask a model to do, and paying frontier prices and accepting network latency for them is poor engineering.
On-device also changes the compliance position. A model that runs locally sends nothing anywhere, which resolves data residency, retention and third-party processing questions by construction rather than by contract. That is worth more to a European or healthcare buyer than several points of benchmark performance.
The category is filling from several directions
Local inference has become a contested area in the space of a fortnight.
Anker released the Eufy MindBase, a home hub with a 26-TOPS chip, 64GB of internal storage and support for up to 48TB externally, running camera and security analysis on-device. The Information reported that OpenAI bought tens of thousands of Macs for reinforcement learning and computer-use training, and that Nvidia views Apple as its principal rival in local AI. Google shipped its third Flash model in six weeks.
The common driver is that the economics of sending everything to a frontier model do not work for high-volume, low-value calls, and the privacy position is worse.
Where the small open models are coming from
OpenBMB is a Chinese open-source group, and it is part of a pattern that has become the defining feature of the open-weight ecosystem this year.
Within the same fortnight, Z.ai released GLM-5.3-Flash as a 320-billion-parameter open-weight multimodal model with 18 billion active parameters at a tenth of its predecessor's cost. Alibaba refreshed Qwen3.8-Max at 2.4 trillion parameters with a one-million-token context window. MBZUAI published six K2 Horizon models from 0.9 billion to 375 billion parameters under Apache 2.0, with weights, code, training data and methods, supported by vLLM, SGLang, Ollama and Unsloth from launch.
Most of the serious open-weight releases now come from Chinese labs and Gulf-funded institutes. That has a strategic dimension the licence does not address: the models are genuinely free to use, and the ecosystem that depends on them depends on decisions made elsewhere.
It also complicates the ownership question just settled above it. Nvidia is buying Hugging Face — the platform through which nearly all of these are distributed — for $12.9 billion.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters