Vision on 3B of active parameters. OpenBMB’s MiniCPM-V-4.7-35B-A3B (October 7, 2026) is a 35.21B-total, 3B-active vision-language model, Apache 2.0, reading screenshots, documents and images through the same interface as text. The official GGUF carries 46,970 downloads in its first week; 1B and DSpark builds shipped alongside.
Fit math and one surprise. 20.2GB of q4 weights on a 262,144-token context, and the smallest KV cache of its class: 40 layers with 2 KV heads at 256 head-dim is just 40MB per 1,000 tokens, the same figure as models a quarter of this size. Screenshots and multi-page documents fill context faster than text, so the generous cache budget matters more for a VLM than a chat model.
Where it sits. Apache 2.0 with every quant official (GGUF, FP8, AWQ, GPTQ, and a DSpark build) puts it next to the Qwen3.8-Omni-Flash tier for local multimodal work, with no license negotiation. A 1B sibling anchors the phone tier.