MiniCPM-V 4.7 offers two model sizes for local vision tasks

OpenBMB released MiniCPM-V 4.7 on October 12 with two downloadable vision-language models: 35B-A3B and 1B. Both have documented routes for asking questions about images and video. The eye-catching phone result belongs specifically to an optimized version of the larger mixture-of-experts model: OpenBMB reports more than 13 decoded tokens per second using 3 GB of memory on an iPhone 17 Pro Max. That is a vendor measurement of one setup, not a hardware requirement or speed forecast for every installation.

The big change

• What changed: OpenBMB has published both a large mixture-of-experts checkpoint and a much smaller model in the same vision series, with example code for image and video prompts and separate quantized downloads. Developers can choose a model and runtime around their actual task and device.
• Why it matters: Local visual analysis can keep the inference workload on hardware the developer controls. The larger model's reported phone performance depends on a particular combination of scheduling, quantization and speculative sampling. Downloading its standard checkpoint alone does not reproduce that result.
• What to watch: The useful decision is whether the chosen runtime can handle the required image detail, video frames and response time on the target device. OpenBMB's model cards expose the weights and processing controls, but its launch numbers do not answer that application-specific question.

What a developer can actually use

The 35B-A3B model card and 1B model card list downloadable checkpoints under Apache 2.0. The larger variant is a mixture-of-experts model; the 1B release is the smaller option. These are alternative sizes for the same broad class of image-and-text tasks. The phone speed claim applies only to 35B-A3B.

OpenBMB's inference guide uses Hugging Face Transformers 5.18 or later with PyTorch, torchvision and a video decoder. Its example loads AutoProcessor and AutoModelForImageTextToText from a named checkpoint. A user message combines an image URL with a text question; the processor formats the message, the model generates tokens, and the example decodes a text answer. The video example replaces the image with a video URL and asks for a description along the timeline. This is visual understanding and text generation, including documents and multiple images in the project's capability list. It is not a video generator or the separate MiniCPM-o speech model.

The guide exposes choices that affect the job. Its default 16x visual downsampling saves tokens, while 4x retains more detail at higher processing cost. Video controls include a 128-frame setting in the example and options for frame sampling and stacking. A developer extracting small text from a document or tracking brief events in a clip should evaluate those settings with representative material. The documentation also warns that torchcodec can conflict with some CUDA versions and describes PyAV or a matching PyTorch/CUDA install as alternatives.

OpenBMB lists SGLang, vLLM, llama.cpp and Ollama support, alongside GGUF and several other quantized variants. The model cards also show a transformers serve example with an OpenAI-compatible chat endpoint for image input. These paths have different packaging and device needs. Neither the repository nor the cards turn the standard Transformers example into the optimized iPhone run automatically. The weights and code are accessible; a working local deployment still requires a compatible runtime, enough storage and memory for the selected variant, and testing against the intended input.

Reading the 3 GB and speed claim

The release guide's on-device section explains the bottleneck behind its phone result. A mixture-of-experts model activates only part of its total weights for each token, yet successive tokens may call on different experts. Keeping weights in storage and loading them as needed reduces resident memory but can stall decoding. OpenBMB says its Lookahead Expert Scheduling predicts which experts to fetch for the next layer, while actual routing still decides which experts are computed. Fine-grained layerwise quantization reduces weight traffic, and DSpark speculative sampling contributes to decoding speed.

OpenBMB attributes the 13-plus tokens per second at 3 GB figure to that combined setup on an iPhone 17 Pro Max. Its accompanying fivefold speedup language and benchmark comparisons are likewise company-run results. The public text does not give an independently reproduced run covering other phones, prompt lengths or video workloads, so those figures cannot be used as a general deployment budget. In particular, 3 GB describes reported memory in that test, not the size of the 35B checkpoint or all memory needed by an application.

For a developer choosing between variants, the smaller checkpoint is the straightforward starting point when device limits dominate; the larger MoE is the option to evaluate when capability is worth a more involved runtime. OpenBMB claims both improve the balance of quality and inference time against selected comparison models in its Instruct evaluation. BIG CHANGE has not tested either model. A useful evaluation would pair the reader's own image or video cases with the chosen quantization, processing settings and target hardware, then measure answer quality, peak memory and time to a usable response. Token decoding rate alone leaves out visual preprocessing and task accuracy.

Sources & further reading

• OpenBMB MiniCPM-V repository, accessed October 13, 2026. Its dated update records the October 12 release, names both variants, distinguishes the MiniCPM-V vision series from MiniCPM-o, and states the code and weights license. The phone figure and performance comparisons are OpenBMB's claims.
• OpenBMB MiniCPM-V 4.7 guide, accessed October 13, 2026. The first-party guide documents model loading, image and video message formats, visual processing controls, decoder dependencies and the techniques OpenBMB says produced its phone result. These instructions were reviewed, not run by BIG CHANGE.
• 35B-A3B model card and 1B model card, accessed October 13, 2026. The official checkpoint pages identify the Apache 2.0 license, downloadable models and quantized variants, and show supported access examples. Availability of weights does not establish performance on a particular device.