A dazzling desktop pet, the world’s best model under 2B, an open dataset… MiniCPM5-1B is officially released and open-sourced!

2026-05-26

From May 25 to 29, ModelBest and OpenBMB are co-hosting the On-Device Large Model Open Source Week, unveiling one killer capability of on-device models each day. The peak of on-device models is not just the tip of the iceberg, but the entire iceberg. Today brings the second installment: MiniCPM5-1B, the small but mighty on-device text model.

Today, ModelBest, together with Tsinghua University and the OpenBMB open-source community, officially releases and open-sources MiniCPM5-1B, the latest generation of its on-device text foundation model.

MiniCPM5-1B once again raises the ceiling on intelligence density: with only 1B parameters, it surpasses every model under 2B on the internationally recognized AA-Index leaderboard. Compared with Qwen3.5-2B, released three months earlier, MiniCPM5-1B is not only more capable but uses half the parameters.

This further confirms the density law we have observed consistently: the intelligence density of large models keeps rising, doubling roughly every 3.5 months. Smaller models are carrying higher intelligence density.

Results across multiple public leaderboards show that MiniCPM5-1B ranks best among base models under 2B parameters worldwide in overall text performance:

Across knowledge, mathematical reasoning, code reasoning, and tool calling, MiniCPM5-1B surpasses base models of the same size, including Qwen3.5-0.8B and LFM2.5-1.2B-Thinking.

On the authoritative Artificial Analysis (AA) leaderboard, MiniCPM5-1B scores 17.9, ranking first on the "small models" list and even beating Qwen3.5-2B (16.3) to become the strongest open-source base model under 2B parameters worldwide!

AA leaderboard results

It is also on-device friendly, with an extremely low barrier to deployment and operation: after INT4 quantization, the weights are only 0.5GB and can run on phones, in browsers, and on almost any hardware within reach.

Even at 1B parameters, the capable MiniCPM5-1B can drive plenty of fun and impressive applications — such as a "desktop pet" that runs on everyone's device.

This article includes video content. Please view it in the original WeChat post.

What is more unusual is that its Base Model version was pre-trained by ForgeTrain, an AI training framework developed in-house at ModelBest.

ForgeTrain is the world's first production-grade large-model pre-training framework written entirely by AI, with zero human programmers involved in writing framework code, and it trains 10% faster than NVIDIA's Megatron.

A framework forged by AI itself has trained the world's strongest text base model at the 2B scale — showing that recursive self-improvement (RSI) through "AI making AI" is no fantasy but a reality already unfolding, and may arrive even earlier than the 2028 that Anthropic AI founder and CEO Dario Amodei predicted.

MiniCPM5-1B is now fully open source, including model weights, training datasets, and deployment solutions. Here is the full introduction to this small but mighty model.

1B parameters, powering a dazzling desktop pet

MiniCPM5-1B's overall performance further changes what the industry expects from on-device models: a base model at around one billion parameters (1B) can drive many powerful and impressive applications.

For example, a desktop pet anyone can afford, running on your phone or computer, keeping you company anytime, anywhere, online or offline:

➤ Project page

🔗 https://github.com/OpenBMB/MiniCPM-Desk-Pet (this project is a derivative of clawd-on-desk)

With continued advances in training techniques and data engineering, the capability frontier of 1B models has been pushed much further.

Its deployment cost is almost negligible: no GPU cluster, no cloud API. An ordinary laptop, a phone, or even a browser tab is enough.

If large models with tens of billions of parameters are giants that need a data center to support them, MiniCPM5-1B is an AI "desktop pet" anyone can afford — it waits in a device right beside you, always ready, needing only a little compute and a few tokens to "feed" it, running even without a network, and far smarter than you might expect.

This is what is truly exciting about on-device AI: it is not about discounting a cloud model's capability to cram it into a small device, but about making a small model built for on-device environments strong enough on its own to drive real applications.

Evaluations on public datasets show that MiniCPM5-1B achieves the best overall performance at its parameter scale, surpassing same-size models (including Qwen3.5-0.8B and LFM2.5-1.2B-Thinking) across general knowledge, mathematical reasoning, code reasoning, and tool calling:

This is not a narrow lead on a single metric but a comprehensive, systematic lead.

On the AA leaderboard, MiniCPM5-1B's score of 17.9 even beats base models with twice the parameters, including the Qwen3.5-2B thinking model (16.3), making it the best open-source base model under 2B parameters worldwide in overall capability:

Moreover, MiniCPM5-1B is small in size but extremely high in intelligence density:

High intelligence density: data sets the ceiling

For a 1B model to reach excellent performance, training data quality matters more than parameter count.

During training, MiniCPM5-1B used a tiered data governance system that classifies pre-training data into five quality tiers, L0 through L4, each with its own cleaning, filtering, and quality-control standards.

On that basis, the research team carried out large-scale synthesis of high-quality pre-training data in three key corpus directions:

High-knowledge-density Chinese web corpus

High-knowledge-density English web corpus

High-quality synthetic math corpus

ModelBest's core philosophy on data governance: rather than flooding a model with massive amounts of low-quality data, nurture it with carefully selected high-density data. At 1B parameters, the quality of every training example directly affects final performance — one of the key reasons MiniCPM5-1B can comprehensively surpass models of its size.

The high-quality synthetic dataset Ultra-FineWeb-L3 will be open-sourced alongside the model for community use and research.

For details on the data governance methodology, see the technical report: 🔗 arxiv.org/pdf/2602.09003

AI makes AI

The Base Model version of MiniCPM5-1B was pre-trained on Huawei Ascend by ForgeTrain, the AI training framework we are about to release.

ForgeTrain is the world's first production-grade large-model training framework written entirely by AI — all code generated by AI with zero code from human engineers. On NVIDIA H100 it trains faster than NVIDIA's own Megatron framework, 10% faster, equivalent to a 10% drop in training cost.

MiniCPM5-1B not only reaches a very high level of intelligence in a small package; it further validates the "AI making AI" path: a training framework written entirely by AI can train the world's best base model at its size.

"AI making AI" is no longer an academic direction that still needs to be argued; it is already producing real, verifiable, performance-leading results. ForgeTrain + MiniCPM5-1B is the latest data point.

On-device ubiquity: 90%+ of devices can run it

There is only one question that matters for an on-device model: can the device in the user's hand run it?

MiniCPM5-1B's answer: almost any of them.

0.5GB — that means MiniCPM5-1B, a highly capable text large model, takes up less space than a short video.

More importantly, it is broadly compatible with runtime environments:

Have a GPU? Run FP16 directly and get full performance.

Only a CPU? ModelBest recently open-sourced ArcLight (github.com/OpenBMB/ArcLight), its in-house CPU inference framework, together with Tsinghua University and the OpenBMB community. ArcLight is deeply optimized for pure CPU environments, so conversations stay smooth even without a GPU.

Don't even want to set up a local environment? MiniCPM5-1B can run directly in a browser: open a web page and it works, with zero installation and zero configuration.

We want MiniCPM5-1B to be a truly zero-barrier on-device model — it doesn't care about your hardware, your platform, or who you are. If you want to run it, you can.

One line of code to wake up your AI desktop pet

However good a model is, developers won't use it if deployment is hard and the ecosystem is incompatible. MiniCPM5-1B was designed around developer friendliness from day one:

Full coverage of mainstream fine-tuning frameworks: Llama_factory and ms_swift

Inference framework support: SGLang, vLLM, llama.cpp, ollama, Hugging Face, and ArcLight (our in-house CPU inference framework)

MiniCPM5-1B also ships with companion installation skills. If you use Claude Code or a similar AI coding tool, you can have it handle the environment setup and deployment of MiniCPM5-1B for you — using AI to install AI, without typing a single command yourself.

Your next desktop pet is only 0.5GB. Come adopt one.

MiniCPM5-1B is now fully open source:

➤ Hugging Face

🔗 https://huggingface.openbmb.com/model/openbmb/MiniCPM5-1B

➤ GitHub

🔗 https://github.com/OpenBMB/MiniCPM

➤ ModelScope

🔗 https://modelscope.cn/models/OpenBMB/MiniCPM5-1B

➤ GitCode

🔗 https://ai.gitcode.com/OpenBMB/MiniCPM5-1B

➤ Modelers

🔗 https://modelers.cn/models/OpenBMB/MiniCPM5-1B

If you would like to join ModelBest's MiniCPM team, send your resume to career@modelbest.cn.

Open roles:

Language large-model algorithm researcher: take part in research and development across the full large-language-model pipeline, including but not limited to data training, architecture innovation, and post-training.

Language large-model AI Infra: take part in the design, development, and optimization of training and inference infrastructure, supporting large-scale model training.