AnswersPublished24 sources
MiniCPM-V 4.7: A 35B MoE Model Built for On-Device Multimodal AI
MiniCPM V 4.7 offers 35B A3B and 1B multimodal models; its first in series MoE activates about 3B parameters per token. OpenBMB’s changelog dates the open source release to October 12, 2026—not October 14—and lists Apache 2.0 licensing for the code and weights.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- MiniCPM V 4.7 offers 35B A3B and 1B multimodal models; its first in series MoE activates about 3B parameters per token.
- OpenBMB’s changelog dates the open source release to October 12, 2026—not October 14—and lists Apache 2.0 licensing for the code and weights.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did ModelBest’s October 14, 2026 open-source release of MiniCPM-V 4.7 introduce, including its 35B-A3B and 1B variants, the series’ fir. Article summary: MiniCPM-V 4.7 introduced a two-model, on-device multimodal family, led by the series’ first mixture-of-experts (MoE) model, with an emphasis on local image, document and video understanding.[1][2][10] One date correction. Topic tags: general, education, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, cha
MiniCPM-V 4.7 is an open-source multimodal model family designed for local use, with two options: a 35B-A3B mixture-of-experts (MoE) model and a smaller 1B model. The flagship is the first MoE in the MiniCPM-V series. OpenBMB reports that it can decode at more than 13 tokens per second using about 3GB of memory on an iPhone 17 Pro Max—but that figure is a company-reported result, not an independent test across devices.9
13
What MiniCPM-V 4.7 includes
The 35B-A3B model has about 35 billion total parameters, with about 3 billion active per token. The 1B variant provides a smaller alternative. Both build on the series’ image and video understanding capabilities.2
31
32
OpenBMB’s project materials list the model code and weights under the Apache-2.0 license and provide code and documentation through the MiniCPM-V GitHub repository.9
14
How the on-device speed claim works
OpenBMB attributes the flagship’s reported phone performance to Lookahead Expert Scheduling (LES), used alongside fine-grained layer-wise quantization and DSpark speculative sampling. LES anticipates which experts a following layer may need so their weights can be loaded earlier.1
13
The headline figure is a decoding speed—the rate at which the model generates output—not a measure of total time from opening an app or submitting an image to receiving a complete answer. The reported result is specific to an iPhone 17 Pro Max and should not be taken as a guarantee of the same speed or memory use on every supported device.13
What the published comparisons show
OpenBMB’s documentation reports that, in Instruct mode, 35B-A3B achieves a higher overall score in less inference time than Qwen3.6 35B-A3B. It also says the 1B model improves on MiniCPM-V 4.6 and outperforms Qwen3.5 0.8B in overall score and inference efficiency.10
14
For OCR, OpenBMB reports an OCRBench score of 90.4 with 16× visual-token compression.12 Benchmark scores depend on the test setup, so results from different compression settings should not be treated as directly comparable. These are developer-reported evaluations, not independent confirmation of performance on every task or device.
12
Platforms and developer tools
The project advertises support for iOS, Android and HarmonyOS, as well as chip platforms from Intel, AMD, Arm, Rockchip and Houmo. These compatibility claims describe the project’s stated coverage; they do not establish identical performance across all hardware.30
37
For developers, the release materials list inference options including SGLang, vLLM, llama.cpp and Ollama, and fine-tuning tools including SWIFT and LLaMA-Factory. They also describe Self-Replay Fine-Tuning (SRFT) as a recipe for extending model capabilities at low cost while reducing forgetting of existing capabilities.20
Release date: October 12, not October 14
Some coverage is dated October 14, 2026, but OpenBMB’s project changelog records the open-source release on October 12, 2026.8
19 For the release date, the changelog is the more direct source.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "MiniCPM-V 4.7: A 35B MoE Model Built for On-Device Multimodal AI"?
MiniCPM V 4.7 offers 35B A3B and 1B multimodal models; its first in series MoE activates about 3B parameters per token.
What are the key points to validate first?
MiniCPM V 4.7 offers 35B A3B and 1B multimodal models; its first in series MoE activates about 3B parameters per token. OpenBMB’s changelog dates the open source release to October 12, 2026—not October 14—and lists Apache 2.0 licensing for the code and weights.