OpenBMB releases MiniCPM-V 4.7
An open (Apache 2.0) 35B mixture-of-experts vision-language model with ~3B active parameters, said to run on an iPhone in 3 GB of memory
Partly confirmed
The takeaway
The model card claims 13+ tokens per second in 3 GB of memory on an iPhone 17 Pro Max and a higher overall score in less time than Qwen3.6 35B-A3B.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 13 October 2026
Your AI and this story
- GPT-6 Astra160 days after its cutoff
- Claude Opus 5.599 days after its cutoff
- Gemini 4 Argon190 days after its cutoff
- Grok 4.7129 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 99 days before it.
Key facts
- Repositories (Hugging Face API creation times): GGUF builds of both sizes on Oct 6, 2026; openbmb/MiniCPM-V-4.7-35B-A3B on Oct 7; FP8 and DSpark draft models on Oct 8; int4, AWQ and GPTQ quantizations on Oct 6-9. No blog post or paper was found; the date here is that of the main repository
- Architecture (model card): ‘the first MoE model in the MiniCPM-V series’; 35B total parameters, about 3B active; language backbone credited to Qwen3.5 and Qwen3.6, vision encoder SigLIP2, with ‘the intra-ViT early compression technique from LLaVA-UHD v4’ and ‘mixed 4x/16x visual token compression’
- Inputs: single images, multiple images and video; output: text. License: Apache 2.0
- On-device claim (vendor): ‘a decoding speed of 13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max’, using Lookahead Expert Scheduling (loading only the experts about to be needed), layer-wise quantization and DSpark speculative sampling with a small draft model
- Platforms named: iOS, Android, HarmonyOS, and ‘day-0 support for chip platforms from Rockchip, Houmo, Intel, AMD, and Arm’
- Benchmark claim (vendor, no numbers in text): in Instruct mode the 35B-A3B model ‘achieves a higher overall score in less inference time than Qwen3.6 35B-A3B’
- Runs with transformers (>=5.18), vLLM, SGLang, llama.cpp (GGUF) and Ollama (openbmb/minicpm-v4.7). Not served by any Hugging Face inference provider; no API price
- Uptake by Oct 13 (Hugging Face API): the 35B-A3B GGUF repository about 10,100 downloads, the 1B GGUF about 2,100; the main repository 448 likes
What happened
OpenBMB’s MiniCPM-V line is a family of small open vision-language models built to run on phones. Version 4.7 is its first mixture-of-experts model: 35B parameters are stored, but only about 3B are active for each token, and a scheduling technique loads only the experts that are about to be used. That is how OpenBMB explains the claim of running a 35B model in 3 GB of phone memory. A 1B version and several quantized builds were published alongside.
We read the Hugging Face model card and the Hugging Face API listing. The benchmark comparison with Qwen3.6 35B-A3B is given only as charts, there is no paper, and we found no independent measurement of the iPhone claim, so confidence is medium.
Why it matters
If the memory claim holds, it is a notable step for on-device multimodal models: a model of the size class that normally needs a workstation GPU answering questions about images and video on a phone, under a permissive license.
Sources
5 sources from 3 sites. Numbers match the chips in the text.
5 sources: 5 primary
Primary
- Hugging Face: openbmb/MiniCPM-V-4.7-35B-A3B (model card)huggingface.co, official
- Hugging Face: OpenBMB organization (all MiniCPM-V 4.7 variants)huggingface.co, official
- GitHub: OpenBMB/MiniCPM-Vgithub.com, code
- GitHub: MiniCPM-V 4.7 documentationgithub.com, docs
- Ollama: openbmb/minicpm-v4.7ollama.com, code
Changes
- Filed (primary source is the Hugging Face model card)