MiniCPM-V 4.7 35B-A3B

Open-weight multimodal MoE, first MoE in the MiniCPM-V series. HF createdAt 2026-10-07, inside the watcher 7-day window as of 2026-10-17. text_config model_type qwen3_5_moe_text: LLM geometry matches Qwen3.5-35B-A3B (hidden 2048, 40 layers, 16 Q / 2 KV heads, head_dim 256, 256 experts, 8 per token, shared expert intermediate 512, moe intermediate 512). layer_types is 30 linear_attention + 10 full_attention, full_attention_interval 4, full attention at indices 3, 7, 11, 15, 19, 23, 27, 31, 35, 39. partial_rotary_factor 0.25, rope_theta 1e7, mrope interleaved (mrope_section [11, 11, 10]). rope_scaling is absent, so context_scaling is null and context_scaling_method is mrope. vocab 248144, mtp_num_hidden_layers 1, linear_conv_kernel_dim 4, linear_num_key_heads 16, linear_num_value_heads 32, linear key/value head dim 128. KV CACHE: 10 full-attention layers x 2 kv heads x head_dim 256 x 2 x 2 bytes = 20,480 B/token. The 30 linear-attention layers hold a fixed recurrent state, so this is the growing cache only. FLOPS: flops_per_token_active = 2 x 3e9 = 6e9 uses the vendor 3B activated figure; the 3B is the 35B-A3B nameplate, not a count derived here. safetensors total is 35.213B BF16, close to the 35B nameplate. Vision: SigLIP-style encoder, 27 layers, hidden 1152, 16 heads, patch 14, image_size 980, intermediate 4304, gelu_pytorch_tanh, insert_layer_id 6 (vision features injected at LLM layer 6), downsample_mode 16x. The card claims mixed 4x/16x visual token compression via LLaVA-UHD v4 intra-ViT early compression. ON-DEVICE (vendor_claim): the card states 13+ tokens/s decode with 3 GB of memory on iPhone 17 Pro Max, via Lookahead Expert Scheduling (prefetch only; token-level routing unchanged), fine-grained layer-wise quantization, and DSpark speculative sampling. The card also claims a higher overall Instruct-mode score in less inference time than Qwen3.6-35B-A3B; those numbers are in images, not in the card text, so they are not copied here. Apache-2.0. The card says it serves in SGLang, vLLM, llama.cpp, and Ollama.