35B-A3B: 35.21B stored / approximately 3B active (MoE); 1B: 1.55B stored (1.3B-class dense)
262,144 tokens
Apache 2.0
Same-day open-weight release of the MiniCPM-V 4.7 multimodal family in 35B-A3B and 1B variants, supporting text, single-image, multi-image, document, and video inputs with textual responses. The flagship checkpoint contains exactly 35,212,875,824 BF16 parameters and uses the MiniCPM-V family's first Mixture-of-Experts language backbone, derived from Qwen3.6. Its 40-layer hybrid stack repeats three linear-attention layers followed by one full-attention layer, contains 256 routed experts with eight selected per token plus a shared expert, and includes a one-layer multi-token-prediction module. The compact variant contains exactly 1,554,476,272 BF16 parameters, uses a dense 24-layer Qwen3.5-derived language backbone, and follows the same three-to-one linear/full-attention pattern. Both variants pair their language model with a 27-layer SigLIP2-derived vision encoder, support mixed 4x and 16x visual-token compression, and use canvas M-RoPE to preserve two-dimensional relationships among image slices and temporal order among video frames. OpenBMB positions both models on a new multimodal performance-efficiency Pareto frontier. Its instruct-mode evaluation reports 91.72 on MMBench, 80.22 on MMMU, 91.20 on OCRBench, 94.24 on OmniDocBench v1.6, 76.40 on Video-MME without subtitles, and 68.90 on LongVideoBench for the 35B-A3B model at 4x visual compression. Reported language scores for that checkpoint include 90.21 on CMMLU, 76.46 on AIME26@avg16, 56.57 on LiveCodeBench v6, and 84.65 on tau2-Bench. The 1B model reports 83.17 on MMBench, 86.10 on OCRBench, 68.05 on ScreenSpot, and 54.50 on LongVideoBench, with OpenBMB claiming that it substantially exceeds Qwen3.5-0.8B in both overall score and inference efficiency. For local MoE deployment, OpenBMB introduces Lookahead Expert Scheduling, which predicts cross-layer routing only for expert prefetch and scheduling while preserving the model's original token-level routing. Combined with fine-grained layer-wise quantization and DSpark speculative sampling, OpenBMB reports 13-plus decoding tokens per second using approximately 3 GB of working memory on an iPhone 17 Pro Max. The release supports iOS, Android, HarmonyOS, Intel, AMD, Arm, Rockchip, and Houmo hardware, with deployment through Transformers, SGLang, vLLM, llama.cpp, Ollama, and multiple official BF16, FP8, int4, AWQ, GPTQ, GGUF, and DSpark artifacts. OpenBMB also released Self-Replay Fine-Tuning for extending model capabilities while reducing forgetting.