OpenBMB
An open-source large model research team focused on large language models and multi-agent systems research, and the developer of open-source projects including ChatDev.
Timeline (last 90 days)
OpenBMB is a Tsinghua-affiliated open-source team with mature experience in model open-sourcing and community operations.
The OpenBMB team released the open-source TTS project VoxCPM2, which has gained nearly 40,000 stars on GitHub.
The implementation, including expert scheduling and DSpark integration, has been open-sourced in llama.cpp-omni.
Lookahead Expert Scheduling (LES) technology, without changing the model's actual routing logic, hides the bottleneck of moving expert weights from flash storage to computation by predicting the experts likely needed by the next layer and preloading their weights.
Engineering practices publicly disclosed by OpenBMB enable the 35B-A3B MoE model MiniCPM-V 4.7 to run on an iPhone at 20.5 tok/s without skipping any experts.
OpenBMB's MiniCPM-V 4.7 can run a 35B vision MoE on an iPhone with about 3GB.
OpenBMB's MiniCPM-V 4.7 continues the combination of a SigLIP vision encoder plus a Qwen3.5 language model backbone, supports two visual downsampling modes, 4x and 16x, and adds canvas M-RoPE to lay images on a 2-D canvas to assign (temporal, height, width) positions to visual tokens.
ModelBest (OpenBMB) released MiniCPM-V 4.7, the newest member of the MiniCPM-V series, including 35B-A3B, the series' first MoE architecture variant.
OpenBMB used reinforcement learning combined with OPD (Omni-Policy Distillation) to merge the capabilities of 16 expert models (including 5 agent experts) into MiniCPM5-2B.
MiniCPM5-2B is a 2-billion-parameter model launched by OpenBMB.
4 more timeline events
All Facts (14)
OpenBMB is a Tsinghua-affiliated open-source team with mature experience in model open-sourcing and community operations.
50%UnverifiedThe OpenBMB team released the open-source TTS project VoxCPM2, which has gained nearly 40,000 stars on GitHub.
50%UnverifiedLookahead Expert Scheduling (LES) technology, without changing the model's actual routing logic, hides the bottleneck of moving expert weights from flash storage to computation by predicting the experts likely needed by the next layer and preloading their weights.
50%UnverifiedEngineering practices publicly disclosed by OpenBMB enable the 35B-A3B MoE model MiniCPM-V 4.7 to run on an iPhone at 20.5 tok/s without skipping any experts.
50%UnverifiedOpenBMB's MiniCPM-V 4.7 can run a 35B vision MoE on an iPhone with about 3GB.
50%UnverifiedOpenBMB's MiniCPM-V 4.7 continues the combination of a SigLIP vision encoder plus a Qwen3.5 language model backbone, supports two visual downsampling modes, 4x and 16x, and adds canvas M-RoPE to lay images on a 2-D canvas to assign (temporal, height, width) positions to visual tokens.
50%UnverifiedOpenBMB used reinforcement learning combined with OPD (Omni-Policy Distillation) to merge the capabilities of 16 expert models (including 5 agent experts) into MiniCPM5-2B.
50%UnverifiedMiniCPM5-2B is a 2-billion-parameter model launched by OpenBMB.
50%UnverifiedChatDev is a multi-agent collaboration framework open-sourced by the OpenBMB team, with the goal of completing the entire software development process through multiple AI agents driven by large language models.
50%UnverifiedThe MiniCPM-V series has demonstrated the potential of small models in the field of multimodal understanding.
50%UnverifiedThe implementation, including expert scheduling and DSpark integration, has been open-sourced in llama.cpp-omni.
50%UnverifiedModelBest (OpenBMB) released MiniCPM-V 4.7, the newest member of the MiniCPM-V series, including 35B-A3B, the series' first MoE architecture variant.
50%UnverifiedThe OpenBMB team released MiniCPM5-2B, a small language model with only 2B parameters.
50%UnverifiedOpenBMB发布了面向LLM预训练的开放英文网页语料库UltraFineWeb，总量超过1万亿token、约11.4亿篇文档，全部来自Common Crawl并以Apache 2.0许可证开放
50%