MiniCPM-V 4.7 Went Up Oct. 6 With No Model Card; Its 3 GB iPhone Claim Still Lacks a Checkpoint
A.I. / news
MiniCPM-V 4.7 Went Up Oct. 6 With No Model Card; Its 3 GB iPhone Claim Still Lacks a Checkpoint
OpenBMB's 35.2B-parameter mixture-of-experts vision model now carries an Apache-2.0 tag. The card publishes no scores in text and does not say which weights run in 3 GB.
OpenBMB's MiniCPM-V 4.7 35B-A3B, a 35.2-billion-parameter vision-language model, appeared on Hugging Face on the evening of Oct. 6 with no model card and no declared licence. The page as of Saturday, Oct. 17 tags it Apache-2.0.
The repository openbmb/MiniCPM-V-4.7-35B-A3B was created at 18:36 UTC on Oct. 6 and last committed 12 minutes later, according to an OrcaRouter post published that day. It had 3 likes and 0 downloads, and a direct fetch of the README returned HTTP 404.
What the card says now
The current model card calls the model the first mixture-of-experts release in the MiniCPM-V series. The language backbone comes from Qwen3.5 and Qwen3.6 and the vision module is SigLIP2, per its acknowledgements. Weights are BF16 safetensors and the files are not gated.
The licence is a tag, Apache-2.0, with no licence text or commercial terms on the page. The card gives no release date and no context length.
Third-party figures fill those gaps. DataLearner lists a release date of Oct. 12, 256 experts with 8 chosen per token, about 3 billion active parameters and a 262,144-token position limit. It treats that limit as a configuration ceiling, not proof of long-context accuracy.
| Date | Source | What it records |
|---|---|---|
| Oct. 6 | OrcaRouter | Repo created, no README, no licence |
| Oct. 12 | DataLearner | Release date, Apache 2.0 for code and weights |
| Oct. 17 | Hugging Face card | Apache-2.0 tag, charts, no numeric scores |
The 3 GB claim
The card says the model decodes at "13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max". It does not say which checkpoint or quantisation produced that number.
OrcaRouter counts 16 safetensors shards totalling about 70.4 GB for the BF16 weights. Arithmetic on 35.2 billion parameters puts a 4-bit file near 17.6 GB, so the 3 GB figure cannot describe all of the weights held in memory on the phone. Whether it reflects streamed experts, a smaller companion model or something else is not given.
The card's one comparative claim is that in Instruct mode the model "achieves a higher overall score in less inference time than Qwen3.6 35B-A3B". The page shows charts and no numbers in text, and does not say who ran the evaluation.
Benchmark numbers, from a third party
DataLearner reports figures it attributes to OpenBMB, not independently verified. At 4x visual-token compression it lists MMStar at 80.67, OCRBench at 91.20, MMMU at 80.22 and MathVista at 88.00. It ranks the model first of six on MMStar and OCRBench, 16th of 73 on MMMU and 20th of 23 on CharXiv RQ at 75.60.
Switching to 16x compression moves MMStar to 80.13, MMMU to 80.00 and MathVista to 87.30.
- MMStar, 4x80.67 points
- MMStar, 16x80.13 points
- MMMU, 4x80.22 points
- MMMU, 16x80 points
- MathVista, 4x88 points
- MathVista, 16x87.3 points
Source: DataLearner, figures attributed to OpenBMB, accessed 2026-10-17
A drop of 0.54 points on MMStar for a fourfold cut in visual tokens is the most useful number on offer, and it comes from a third-party listing, not the card.
What is still open
OpenBMB has not published a technical report or launch post that the pages reviewed link to. The card does not name a training-data source or give a training cost.
For comparison with other open-weight releases, see Reflection's 501B Apache-licensed weights and the PS5 port that runs a 27B Qwen model. The Apache-2.0 tag is the only licence statement for MiniCPM-V 4.7 so far. OpenBMB has not said when a full card with scores will follow.
Sources
More in A.I.
- 01Qwen3-4B-Deslop Cuts Its Slop Score From 121 to 12.5 on a MacBook Air, and Two of 40 Answers Came Out DamagedThe Show HN fine-tune by Erik Markus Kannike uses a rank-16 LoRA and a 175-byte neuron edit; its blog gives no licence, its Hugging Face card says Apache-2.0.
- 02HybridAgent Paper Claims 10.4% Gain and 2.7x to 3.7x Lower Cost, but Its Own Text Gives 10.4% Two Different MeaningsThe Stanford, UIUC and Together AI preprint ships MIT-licensed 8B and 14B agents, and no independent run of its numbers has been reported.
- 03Gemini 4 Argon Reaches the API at $2 and $10 Per Million Tokens, With a No. 1 Ranking and No ScoreGoogle's October 13 post says Argon tops the Vals Index but publishes no figure for it, and the introductory price doubles on November 30.
- 04DeepSeek Swapped V4.1-Flash-1017 Behind Its API Name, and the New Card Lists NL2Repo at 65.4 Where the Old One Said 64.0The MIT-licensed update calls itself minor and is now what deepseek-flash serves, with no active-parameter count and no release date on the card.