From May 25th to 29th, Facewall Intelligence and OpenBMB jointly hosted the "Edge-Side Large Model Open Source Week," unlocking the killer features of a large edge-side model each day.
From May 25th to 29th, Facewall Intelligence and OpenBMB jointly hosted the "Edge-Side Large Model Open Source Week," unlocking the killer features of a large edge-side model each day. TMiniCPM5-1B Open Source! An AI desktop pet that runs on your phone is here!he pinnacle of large edge-side models isn't just the tip of the iceberg, but the entire iceberg. Today is the second installment of Open Source Week: the edge-side text powerhouse – MiniCPM5-1B, officially released and open-sourced by Facewall Intelligence in collaboration with Tsinghua University and the OpenBMB open-source community, representing the latest generation of edge-side text foundation large models. MiniCPM5-1B once again refreshes the upper limit of model intelligence density: with only 1B parameters, it surpasses all models with fewer than 2B parameters on the internationally renowned AA-Index leaderboard. Compared to Qwen3.5-2B released three months ago, MiniCPM5-1B not only performs better but also reduces the number of parameters by half. This result further validates our continuously observed density law: the intelligence density of large models is continuously increasing at a rate of approximately doubling every 3.5 months. Smaller models are carrying higher intelligence density. According to multiple publicly available benchmarks, MiniCPM5-1B boasts the best overall text performance among base models with a parameter size of 2B or less globally. In dimensions such as knowledge, mathematical reasoning, code reasoning, and tool invocation, MiniCPM5-1B comprehensively surpasses base models of similar size, including Qwen3.5-0.8B and LFM2.5-1.2B-Thinking. On the authoritative Artificial Analysis (AA) benchmark, MiniCPM5-1B's score (17.9 points) ranks first in the "Small Models" category, even surpassing Qwen3.5-2B (16.3 points), making it the strongest open-source base model with a parameter size of 2B globally!
Furthermore, it is device-friendly, with extremely low deployment and operation barriers. After INT4 quantization, the weight is only 0.5GB, which means it can run on mobile phones, in browsers, and even on almost any hardware at your fingertips. Even with only 1B parameters, the powerful MiniCPM5-1B can still drive a large number of interesting and cool applications, such as "desktop pets" running on everyone's devices.
What's even more remarkable is that its Base Model version was pre-trained using ForgeTrain, an AI training framework independently developed by Wallfacer AI. ForgeTrain is the world's first production-grade large-scale model pre-training framework written entirely by AI, with zero human programmer involvement in writing the framework code, and its training speed is 10% faster than NVIDIA Megatron. A framework forged by AI itself trained the world's most powerful text base model at a 2B scale. This demonstrates that "AI creating AI"—Recursive Intelligence (RSI)—is not a pipe dream, but a reality that is already unfolding, and may even be achieved earlier than Anthropic AI founder and CEO Dario Amodei predicted in 2028. MiniCPM5-1B is now fully open source, including model weights, training datasets, and deployment solutions. Below is a complete introduction to this powerful little machine. With 1B parameters, the MiniCPM5-1B, a super-cool "desktop pet," has further enhanced the industry's understanding of edge-side models—a base model with approximately 1 billion parameters (1B) can also drive many powerful and cool applications. For example, an "desktop pet" that everyone can afford, running on your phone or computer, can be a thoughtful companion anytime, anywhere, with or without internet access.
