OpenBMB knowledge base goes live: everything from choosing a model to getting an app running, in one place

2026-07-27

From picking the right model to actually getting it running, developers have more to cross than a single download.

Does the model's capability match the task? Can the device handle it? How do you pin down a deployment problem? Is there already a proven application path? These key answers used to be scattered across model repositories, technical docs, and community discussions.

Now we have brought them together.

The OpenBMB community knowledge base is officially live.

From model information, deployment and fine-tuning, and VRAM and performance to developer practice and frequently asked questions, it gives every exploration a clearer starting point.

Start by getting to know the models and find your MiniCPM

MiniCPM now forms a model family spanning text, multimodal, omni-modal, and speech. Faced with different parameter sizes, input modalities, and deployment methods, new developers often need to answer not "how do I run it" but "which one should I choose."

Under Community Technical Resources, we have organized clear, immediately usable guides for each MiniCPM model and product line, covering:

MiniCPM text models

MiniCPM-V multimodal models

MiniCPM-o omni-modal models

VoxCPM speech models

PilotDeck agents

Starting from the task, the device, and the interaction style, you can quickly learn what each model is good at, where it fits for deployment, and what kinds of applications it can build.

It has to run—and the numbers have to add up

Once a model is chosen, whether a device can run it and how far it can go often comes down to VRAM and performance.

The VRAM and Performance section collects quick-reference VRAM tables for each model, device compatibility references, and the relationships among model weights, KV cache, context length, and inference memory.

Here you can quickly work out:

roughly how much VRAM different precisions and quantization versions need;

which model your device can run;

why context length affects VRAM usage;

which parameters to adjust first when you hit an OOM.

You do not have to start from trial and error every time. Work out the resources first, then choose a more suitable deployment plan.

More than tutorials: MiniCPM in the real world

What is really interesting about a model is always what developers do with it.

So the knowledge base will keep collecting application case studies, deployment practices, and engineering experience from the community. For example:

building a replayable, evaluable local control pipeline in Rust for a 1B on-device model;

running local AI on a NAS with an N150 CPU;

fine-tuning MiniCPM5-1B into an on-device model with a specific persona and interaction style, then deploying it to an Android phone.

These pieces are more than showcases. Wherever possible they keep the model choice, technical path, runtime environment, and the pitfalls encountered, so others can actually reproduce the work and move further.

When you get stuck, look here first

The knowledge base already organizes FAQ entry points around model selection, download, deployment, VRAM, quantization, and long context.

If you do not find an answer, you can join the relevant developer group to keep discussing, or leave your question in the knowledge base. We will reply within three business days.

📌 Tip for asking questions:

Please include the model version, runtime environment, GPU and VRAM, launch command, full error output, and minimal reproduction steps. The more complete the information, the easier it is for community members to help you pinpoint the problem quickly.

This knowledge base is yours to improve, too

We hope this knowledge base is built not only by OpenBMB but by every developer who uses MiniCPM.

An application you got running, a deployment note you wrote down, a compatibility issue you solved—even an attempt that did not work—may be exactly the answer another developer is looking for.

That is why, through the MiniCPM Spark Program and the MiniCPM Global Ambassador Program, we are looking for three kinds of collaborators:

technical practitioners building real applications with MiniCPM;

community builders willing to organize meetups, workshops, and hackathons;

mission advocates who believe in "putting large models as close to users as possible" and are willing to connect developers with the open-source community.

The knowledge base will keep updating model information, technical tutorials, performance data, developer case studies, event information, and FAQs. We welcome you to join with code, projects, benchmarks, reproduction records, deployment scripts, and wild ideas.

Start here and get the model actually running

If you are new to MiniCPM, treat it as an onboarding map.

If you are deploying models or building applications, treat it as an engineering handbook you can consult anytime.

If you have already got your own project running, leave your experience here—it can be a spark that lights the way for those who come after.

We love developers.

This is more than a place to store answers. It is also a place to record every act of creation as it happens.

➤ Knowledge base link

🔗 https://modelbest.feishu.cn/wiki/UtWxwcERfiRIpIkBOjuc3h9tn1D