openbmb/MiniCPM-V-4.7-35B-A3B model card and config.json

Card and config.json read 2026-10-17. HF API createdAt 2026-10-07T07:07:30Z, lastModified 2026-10-15T16:33:30Z, license apache-2.0, safetensors total 35,212,875,824 BF16. config.json text_config model_type qwen3_5_moe_text: hidden_size 2048, 40 layers, 30 linear_attention + 10 full_attention (full_attention_interval 4), 16 Q heads, 2 KV heads, head_dim 256, 256 experts, 8 per token, shared_expert_intermediate_size 512, moe_intermediate_size 512, max_position_embeddings 262144, partial_rotary_factor 0.25, rope_theta 1e7, mrope interleaved mrope_section [11, 11, 10], vocab_size 248144, mtp_num_hidden_layers 1, linear_conv_kernel_dim 4, linear_num_key_heads 16, linear_num_value_heads 32, linear head dims 128. No rope_scaling key. Vision minicpmv4_7_vision: 27 layers, hidden 1152, 16 heads, patch 14, image_size 980, intermediate 4304, insert_layer_id 6, downsample_mode 16x. Card says 35B-A3B is the first MoE in the MiniCPM-V series and claims 13+ tokens/s decode at 3 GB on iPhone 17 Pro Max via Lookahead Expert Scheduling, layer-wise quantization, and DSpark. Benchmark charts are images, not a numeric table. Config: https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B/raw/main/config.json.