OpenBMB Releases MiniCPM-V 4.7: On-Device Multimodal MoE Taking on Qwen3.6

OpenBMB released the multimodal model MiniCPM-V 4.7 on October 17, offering a 35B-A3B version and a 1B version, focusing on on-device multimodal capabilities and document understanding. Niels Rogge of Hugging Face posted a series of threads covering the model's official benchmarks, local deployment, and companion tools, with an overall takeaway: it beats Qwen3.6 on document understanding at the same scale, and the barrier to local deployment is remarkably low — well worth attention.

Confirmed

MiniCPM-V 4.7 ships in two versions: 35B-A3B, the family's first MoE model (35B total / 3B active parameters) with 4x visual token compression; and a 1B version taking just 2.9GB of memory, targeting edge devices.
Official comparison against Qwen3.6-35B-A3B (same 35B total / 3B active architecture): scores about 14 points higher on the OmniDocBench document understanding benchmark.
Local performance: Niels Rogge demonstrated running MiniCPM-V 4.7 with llama.cpp, reaching 100 tokens/second of generation on roughly 3GB of memory, with a full tutorial thread attached.
Released alongside it is the DSpark draft head: speculative decoding draft checkpoints trained specifically for the 35B-A3B target model (plus a 1B version). The draft side has only 5 layers and about 389 million parameters, proposing 15 tokens per forward pass for the main model to verify, with a maximum acceptance length of 7.4.
For deployment, Rogge suggests handing OpenBMB's Hugging Face model page straight to Claude or another coding agent and letting it get things running automatically with llama.cpp.

Why it matters

In the 35B-A3B MoE class, MiniCPM-V 4.7's local performance — roughly 3GB of memory at 100 tokens/second — demonstrates a viable path toward on-device multimodal models.
Lightweight draft heads like DSpark (achieving up to 7.4 acceptance length with just 389 million parameters) offer a new paradigm for speeding up small models via speculative decoding.
The official benchmark claims a 14-point edge over Qwen3.6-35B-A3B; if independently verified, it would reshape the competitive landscape for multimodal models at this scale.

2026-10-17 ~ 2026-10-17 · 5 related posts

Primary sources

MiniCPM-V 4.7: MoE model beats Qwen3.6 on doc understanding, 1B runs in 2.9GB RAM — NielsRogge · 2026-10-17 22:09 (UTC+08:00)
MiniCPM-V 4.7 beats Qwen3.6 on doc understanding, +14 pts on OmniDocBench — NielsRogge · 2026-10-17 22:09 (UTC+08:00)
MiniCPM-V 4.7 runs locally at 100 tok/s using only ~3 GB of memory — NielsRogge · 2026-10-17 22:09 (UTC+08:00)

[source] MiniCPM-V 4.7 runs locally at 100 tok/s using only ~3 GB of memory — NielsRogge · 2026-10-17
Deploy MiniCPM-V 4.7 locally by pointing Claude at its Hugging Face page — NielsRogge · 2026-10-17
[source] MiniCPM-V 4.7 beats Qwen3.6 on doc understanding, +14 pts on OmniDocBench — NielsRogge · 2026-10-17
OpenBMB's DSpark draft head hits 7.4 acceptance length for MiniCPM-V 4.7 — NielsRogge · 2026-10-17
[source] MiniCPM-V 4.7: MoE model beats Qwen3.6 on doc understanding, 1B runs in 2.9GB RAM — NielsRogge · 2026-10-17