MiniCPM-V 4.7
OpenBMB's phone-friendly vision-language models: a 35B-A3B mixture of experts and a 1B model, images and video, Apache 2.0.
In short. MiniCPM-V 4.7 is a family of AI models that look at pictures and videos and answer questions about them. The small 1B version is aimed at phones. The 35B version is a mixture of experts that OpenBMB says can run on a high-end iPhone using clever memory tricks. Both use the Apache 2.0 license.
Summary
OpenBMB released MiniCPM-V 4.7 on 12 October 2026 in two sizes: a 35B-A3B mixture of experts, which is the first MoE model in the MiniCPM-V series, and a 1B model. Both read single images, multiple images and video and write text, under the Apache 2.0 license. The 35B-A3B has about 35.2 billion parameters with roughly 3 billion active per token, 40 layers, 256 experts with 8 active, and a 262,144-token context. The card credits Qwen3.5 and Qwen3.6 for the language backbone and SigLIP2 for the vision module.
The main selling point is efficiency. The vision encoder compresses image tokens early, and developers can choose between 4x and 16x visual token compression to trade detail for cost. For the 35B-A3B model, OpenBMB introduced Lookahead Expert Scheduling, which predicts which experts will be needed and preloads them, and says that with layer-wise quantization and speculative decoding it reaches 13 or more tokens per second using about 3 GB of memory on an iPhone 17 Pro Max. That is the developer's claim; the full weights are still about 23 GB at int4 and 71 GB in BF16, so test it on your own device before relying on it.
OpenBMB also says the 35B-A3B scores higher overall than Qwen3.6-35B-A3B in less inference time in Instruct mode, a vendor comparison. The models run on iOS, Android and HarmonyOS, with day-0 support for Rockchip, Houmo, Intel, AMD and Arm chips, and work with SGLang, vLLM, llama.cpp and Ollama. FP8, int4, AWQ, GPTQ and GGUF builds are published.
Good for
- Reading photos, scanned documents, charts and short videos on a phone or laptop (1B)
- A capable open vision-language model on one high-memory GPU or an Apple silicon Mac with enough memory; the int4 build is 23 GB (35B-A3B)
- On-device multimodal apps on iOS, Android and HarmonyOS
- Commercial use under Apache 2.0, with a fine-tuning method (SRFT) designed to limit forgetting
Watch out for
- The "3 GB on an iPhone" figure depends on expert scheduling and quantization; the full 35B-A3B weights are 23 GB at int4
- Performance comparisons with Qwen3.6-35B-A3B are OpenBMB's own
- It outputs text only; audio input is handled by a different model line (MiniCPM-o)
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 1B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 1B BF16 | 1.55B | 3.56 GB | 1.78 GB | 1.07 GB | Laptop or phone |
| 35B-A3B BF16 | 35.2B 3B active | 81 GB | 40 GB | 24 GB | Gaming GPU |
| 35B-A3B int4 INT4 | 35.2B 3B active | 81 GB | 40 GB | 25 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Best open-source multimodal models in 2026 (images, video and audio)
- Best small open models you can run on a laptop
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is MiniCPM-V 4.7?
- MiniCPM-V 4.7 is a family of AI models that look at pictures and videos and answer questions about them. The small 1B version is aimed at phones. The 35B version is a mixture of experts that OpenBMB says can run on a high-end iPhone using clever memory tricks. Both use the Apache 2.0 license.
- Can I use MiniCPM-V 4.7 commercially?
- Yes. MiniCPM-V 4.7 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is MiniCPM-V 4.7 open source?
- MiniCPM-V 4.7 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run MiniCPM-V 4.7 on my own computer?
- Yes. The smallest variant, 1B, needs about 1.07 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was MiniCPM-V 4.7 released?
- MiniCPM-V 4.7 was released by OpenBMB on October 12, 2026.
- How long a context window does MiniCPM-V 4.7 have?
- MiniCPM-V 4.7 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- MiniCPM-V 4.7 35B-A3B model card (Hugging Face) huggingface.co
- MiniCPM-V repository and release notes (GitHub) github.com
- OpenBMB changelog openbmb.cn
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MiniCPM-o 4.5 OpenBMB | 9.37B | 40K | Apache 2.0 (Permissive) | 11 GB Laptop or phone | |
| MiniCPM5-2B OpenBMB | 1.08B – 2.52B | 128K | Apache 2.0 (Permissive) | 0.745 GB Laptop or phone | |
| LFM3 Liquid AI | 1.73B – 2.17B | 256K | LFM Open 1.0 (Conditional) | 1.19 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |