OpenBMB released a 2.5-billion-parameter model for local and on-device use. It also shared much of its training data, but the strongest test results come from the model maker.
New to this? Read it in simple words
- OpenBMB released MiniCPM5-2B, a small AI model made for local devices. It also shared parts of its training data.
- It has about 2.52 billion parameters and a 131,072-token context window.
- Small open models can make it easier to build private tools that run locally.
- The strongest test results come from OpenBMB itself. Teams should test the model on their own tasks and devices.
- Parameter
- A number inside an AI model that it learns during training.
- Token
- A small piece of text, often part of a word. AI use is priced per token.
- Context window
- The amount of text an AI model can work with at one time.
A small model with a long input window
MiniCPM5-2B has about 2.52 billion parameters. Parameters are the learned numbers inside a model. This is small compared with many cloud models. OpenBMB designed it for local assistants, coding tools, and devices with limited computing power.
The model card lists a native context length of 131,072 tokens. This means the model can receive a large amount of text in one request. A long window does not guarantee that the model will remember every detail correctly.
The release works with common tools such as Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, and MLX. Ready-made formats can make testing easier on Linux computers, local servers, and Apple Silicon machines.
MiniCPM5-2B supports 131,072 input tokens. Its maker and an outside test use different score systems.
OpenBMB shared more than model weights
The group released the final model, earlier checkpoints, smaller formats, and parts of the training data. One agent dataset has 500,000 samples. Another has more than 80,000 examples for reinforcement learning in maths, code, knowledge, and long-context work.
This wider release helps researchers study how the model was built. It can also help teams train a version for their own task. Open data does not remove every question about quality, rights, or missing material, but it gives outsiders more to inspect.
The Apache 2.0 licence allows broad use. Teams still need to read the licence and check the data rules for their own product. They also need security tests before an agent can call tools or reach private files.
Separate maker claims from outside tests
OpenBMB reports an average score of 53.9 across its own comparison table. It says the model leads other open models of a similar size. Most results in that table were produced by OpenBMB, so teams should not treat them as a final answer.
Artificial Analysis ran a separate test and gave the model 15 on its current Intelligence Index. It says this is the highest result among open-weight models below four billion total parameters. Its test uses a different set of tasks, so the two scores cannot be compared directly.
The best next step is a small local trial. Teams should measure answer quality, speed, memory use, power use, and tool safety on real work. A compact model is valuable when it fits the device and completes the task, not only when it wins a table.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3
We used the model card and official code repository for size, licence, formats, data, and maker test results. We used Artificial Analysis for a separate score. We keep the two test systems separate because they use different tasks and scales.