MiniCPM-V 4.7: 35B-A3B MoE and 1B vision models, Apache 2.0, GGUF

The news section of OpenBMB's GitHub README dates the release to 2026-10-12. MiniCPM-V 4.7 comes in two sizes: 35B-A3B, the series' first mixture-of-experts (MoE) model, and 1B, a dense model. OpenBMB lists single-image, multi-image and video understanding. On 2026-10-15 it added DSpark speculative-decoding heads in BF16 and GGUF for both sizes. The 35B-A3B `config.json` sets the text model type to `qwen3_5_moe_text`, with 256 experts, 8 active per token, and `max_position_embeddings` of 262,144. The model guide credits Qwen3.5 and Qwen3.6 for the language backbone and SigLIP2 for the vision module. The README says code and weights are released under Apache-2.0, and both sizes carry that tag on Hugging Face. 35B-A3B ships in BF16, FP8, int4, AWQ, GPTQ and GGUF. 1B ships in all of these except FP8. OpenBMB's 35B-A3B GGUF files include IQ3_XS at 14.48 GB, Q4_K_M at 21.17 GB, Q5_K_M at 24.73 GB, Q8_0 at 36.9 GB and F16 at 69.38 GB. The two vision projector (mmproj) files are 0.91 GB and 1.12 GB. The model guide lists SGLang, vLLM, llama.cpp and Ollama, plus day-0 OpenVINO support on Intel. The llama.cpp guide it links says upstream llama.cpp added support in PR #29416, merged 2026-10-10. It warns that GGUFs converted with an older script compute M-RoPE incorrectly. An `openbmb/minicpm-v4.7` Ollama listing exists, but the linked Ollama guide, last updated 2026-10-14, says Ollama cannot load 4.7 GGUFs until its bundled llama.cpp includes that PR. OpenBMB claims 35B-A3B decodes at 13+ tokens/s with 3 GB of memory on an iPhone 17 Pro Max. It credits Lookahead Expert Scheduling (LES), documented in its llama.cpp-omni repository, plus layer-wise quantization and DSpark. The LES document reports 20.51 tokens/s on an iPhone 18 Pro Max (12 GB) and 17.39 tokens/s on an 11-inch iPad Pro M4 (16 GB). We have not benchmarked either model.