MiniCPM-V 4.7
On-device multimodal MoE (35B-A3B + 1B) that runs on a phone with ~3GB of memory.
About
MiniCPM-V 4.7, released in October 2026, is the latest in ModelBest and OpenBMB's line of small multimodal models built to run on the device itself rather than in the cloud. It ships in two flavors: a dense 1B model and a 35B-A3B mixture-of-experts where only about 3B parameters activate per token — which is how the large one fits in roughly 3GB of phone memory. The team credits a lookahead expert-scheduling technique plus quantization for decoding speeds above 13 tokens per second on an iPhone 17 Pro Max, with day-one support from AMD, Apple, Arm, Intel and NVIDIA.
It handles images, multi-image, documents and video, with 4x and 16x visual-token compression so developers can trade detail for speed. On the team's own benchmarks it scores 91.72 on MMBench and 91.2 on OCRBench, and independent hands-on tests showed it doing chart reasoning and paper-to-Markdown conversion fully offline. Weights, inference and fine-tuning code are Apache 2.0 on the OpenBMB GitHub, with llama.cpp-based demos for iOS, Android and HarmonyOS Next.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Provider-reported: 91.72 MMBench, 91.2 OCRBench, 90.21 CMMLU; 90.9 MMBench / 90.4 OCRBench at 16x visual-token compression
- Provider-reported: >13 tokens/sec on iPhone 17 Pro Max, >17 tokens/sec on iPad Pro M4
- No independent benchmark replication published yet
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: Apache 2.0 weights + code (OpenBMB); provider benchmark table (MMBench 91.72, OCRBench 91.2); independent hands-on tests (QbitAI via EastFrontier) confirming offline phone inference. Strengths: fully open license, genuinely novel on-device MoE story, broad chip-vendor support. Docked for: (1) released October 2026 — days old, so adoption and independent evals are minimal; (2) all benchmark and speed figures are provider-reported. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 5.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Released 2026-10-15; 35B-A3B MoE + 1B dense; Apache 2.0 weights + code (OpenBMB)
- · Provider benchmarks: MMBench 91.72, OCRBench 91.2, CMMLU 90.21
- · Independent hands-on tests confirm offline phone inference (EastFrontier/QbitAI)