MiniCPM-V 4.7 Ships 35B Weights Under Apache-2.0, With Its Benchmark Scores Left in Charts

OpenBMB released MiniCPM-V 4.7 on October 12, a 35-billion-parameter vision-language model that activates 3 billion parameters per token and carries an Apache-2.0 label. The lab says it runs at 13 or more tokens a second in about 3 GB of memory on an iPhone 17 Pro Max, and says its Instruct mode beats Qwen3.6 35B-A3B overall. Neither claim comes with a published number in the text of the model card.

What the MiniCPM-V 4.7 card says

The Hugging Face model card calls it the first mixture-of-experts model in the MiniCPM-V series. Visual tokens are compressed early in the vision encoder, using a design based on LLaVA-UHD v4 with mixed 4x and 16x compression.

The language backbone is credited in the acknowledgements to Qwen3.5 and Qwen3.6, with SigLIP2 as the vision module. The card does not name a single base checkpoint.

OpenBMB's GitHub README lists two sizes, 35B-A3B and a 1B model, and dates the release 2026-10-12. Speculative-decoding heads called DSpark followed on 2026-10-15.

Licence and weights

Both the card and the README label the model Apache-2.0, and the README says "Code and model weights are released under Apache-2.0." The weights are on Hugging Face with no gating notice on the page.

The README also links a separate model licence and usage statement. The terms of that document are not shown on either page, so whether it narrows commercial use is unknown from the sources read here.

How it compares with the Qwen model it names

The comparison OpenBMB draws is with Qwen3.6-35B-A3B, which Qwen released in April 2026 under Apache-2.0 with 256 experts, 8 routed plus 1 shared per token.

	MiniCPM-V 4.7	Qwen3.6-35B-A3B
Total parameters	35B	35B (Hub shows 36B)
Active per token	3B	3B
Licence label	Apache-2.0	Apache-2.0
Context window	Not stated	262,144 native

The Qwen card publishes figures such as 73.4 on SWE-bench Verified and 81.7 on MMMU, and warns they came from varied internal harnesses. The MiniCPM card gives charts but, in its text, no scores. OpenBMB's "higher overall score in less inference time" is a vendor claim with no named independent run.

Hardware and quantised builds

The only memory figure is the phone one: 13 or more tokens per second in about 3 GB on an iPhone 17 Pro Max, according to the README. The card gives no GPU VRAM requirement for the full 35B model. It lists FP8, int4, AWQ, GPTQ and GGUF variants, and the Hub shows 56 quantised models derived from it.

Running it through Transformers needs version 5.18.0 or later, and the card recommends Flash Attention 2 for multi-image and video input.

What is missing

There is no paper. The card names no training data. It does not state a context length. OpenBMB's Self-Replay Fine-Tuning post, dated 2026-10-12, evaluates MiniCPM-V-4.5 and MiniCPM-o-4.5 and does not mention 4.7, so it cannot be used to explain how 4.7 was trained.

The model was listed in Hugging Face's trending models on October 20 with 1,072 likes and 76,410 downloads. That counts interest, not deployments.

For open-weight comparisons on this desk, see DeepSeek's V4.1-Flash-1017 swap behind an API name and the Qwen3-4B de-slop experiment. OpenBMB has not said when a technical report will appear.