MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark Tables
A.I. / news
MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark Tables
OpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
OpenBMB has put MiniCPM-V 4.7 35B-A3B on Hugging Face under an Apache 2.0 tag. It is the first mixture-of-experts model in the MiniCPM-V series, with 35 billion parameters in total and about 3 billion active per token, according to the model card.
The card claims a better overall score than Qwen3.6 35B-A3B in less inference time. It gives no number for either model. Hugging Face's trending list showed 965 likes and 57,935 downloads for the repository on Sunday, October 18.
What the model card says
The card describes an image, multi-image and video model, with video capped at 128 frames. Visual tokens are compressed 4x or 16x. The card credits Qwen3.5 and Qwen3.6 for the language backbone and SigLIP2 for the vision module, and does not say which backbone this checkpoint uses.
An aiweekly report dated October 14 adds a 1B companion model, quantised builds in FP8, AWQ, GPTQ and GGUF, and day-zero support for Rockchip, Houmo, Intel, AMD and Arm chips plus iOS, Android and HarmonyOS. It lists three efficiency techniques: Lookahead Expert Scheduling, fine-grained layer-wise quantization and a DSpark draft model for speculative decoding.
| Item | What is stated | Where |
|---|---|---|
| Parameters | 35B total, about 3B active | Model card |
| Licence tag | apache-2.0 | Model card |
| Context length | Not stated | Model card |
| Training data | No statement | Model card |
| Benchmark scores | Charts only, no figures in text | Model card |
| Quantised builds | FP8, AWQ, GPTQ, GGUF | aiweekly |
The iPhone claim
The headline on-device number is 13 or more tokens a second using about 3 GB of memory on an iPhone 17 Pro Max. That is a vendor claim, and no independent run is cited.
The arithmetic is worth checking. Thirty-five billion parameters at 16 bits is roughly 70 GB of weights. At 4 bits it is about 17.5 GB, still several times the 3 GB figure. The card does not say where the remaining experts are held while the phone decodes, so the claim is best read as a statement about active working memory rather than the full model.
What is missing
The card gives no context length, no hardware requirement for the full model and no release date. It makes no statement about training data. The licence appears only as the repository's Apache 2.0 tag, with no separate terms on the page. The page shows no access gate.
Two other open-weight releases this month are worth reading beside it. Meta's Muse Spark Open carries a $20 million threshold and an attribution requirement in its licence, and we covered DeepSeek's V4.1-Flash-1017 for its benchmark and licence terms. A bare Apache 2.0 tag is the simplest of the licence questions, but the card leaves the others unanswered.
OpenBMB has not published scores that a third party could rerun. Until it does, the claim over Qwen3.6 35B-A3B stands as the developer's word.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 03StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.
- 04Statistics and Optimization Make Up 20% of arXiv Theorems but Under 1% of OpenAI's 722 Math PapersA newsletter analysis of OpenAI's October 6 release finds the fields closest to machine learning nearly absent, and OpenAI has not said why.