Latest from OpenBMB: model releases, technical reports, community events and partnerships.
Latest
From an Idea to 19 Projects | Hi Youth Hackathon: Mianbi Intelligence Special Edition Comes to a Successful Close：If you were given 22 hours, an idea, and an open-source Agent operating system, how far could you go?
The FlagOS X MiniCPM Model Inference Throughput Performance Optimization Challenge is Now Open! Two chips go head-to-head—a 10,000 yuan prize pool awaits your challenge!：FlagOS Open Computing Global Competition (Season 2), Track 2: MiniCPM Model Inference Throughput Performance Optimization—Global Registration Is Now Open!
Community Rally | The First Batch of MiniCPM Global Ambassadors Has Arrived!："Bringing large models into every household"—this is not only OpenBMB's community vision but also the direction in which every open-source contributor is moving forward.
Saves GPU memory, computing power, and costs; small models can be used in this way!：Through engineering methods such as model selection, invocation strategy, edge-cloud collaboration, and inference and deployment optimization, a better balance between cost and performance can be achieved while meeting business performance and experience requirements.
MiniCPM5-1B Released: AA-Index Results and Open Datasets：MiniCPM5-1B scores 17.9 on AA-Index V4.0; see the evaluation table for the sub-2B comparison models and their scores.
MiniCPM-V 4.6 Open-Sourced: 1.3B Parameters and Benchmark Results：Supports 4x and 16x visual token compression; against Qwen3.5-0.8B on standard image input, visual encoding FLOPs drop by over 50% and token throughput is about 1.5x that of Qwen3.5-0.8B.
MiniCPM-o 4.5 Technical Report Released: Full-Duplex Multimodal APIs and Real-Time Local Runtime on an RTX 5070：MiniCPM-o 4.5 uses an end-to-end full-duplex omni-modal architecture; measured minimum VRAM was 12 GB on an NVIDIA RTX 5070 with 4-bit quantization (GGUF Q4_K_M).
UltraData Introduces Its Tiered Data Governance Framework to Power AGI with Scientific Data Governance：ModelBest and Tsinghua released the L0-L4 tiered data management system together with UltraData-Math; its L3 refined tier contains 88B tokens across 81.4 million documents.
DeepResearch Goes Local: AgentCPM-Report, an 8B On-Device Writing Agent, Is Now Open Source：AgentCPM-Report brings DeepResearch-style writing capabilities to local execution in an 8B on-device agent.
VoxCPM2 Deployment and Fine-Tuning Guide: Voice Design, Voice Cloning, Audio Continuation：VoxCPM2 deployment and fine-tuning guide: voice design, voice cloning, and audio continuation.