OpenBMB releases MiniCPM-V 4.7 1B, an on-device vision model with 4x/16x visual token compression

OpenBMB has released MiniCPM-V 4.7 1B, a 1B-variant multimodal model for single-image, multi-image and video understanding, built to run on phones, tablets and local computing devices. The weights are on Hugging Face under openbmb/MiniCPM-V-4.7-1B. According to OpenBMB, the model substantially outperforms Qwen3.5 0.8B in both overall score and inference efficiency in Instruct mode.

Key specifications

Item	Detail
Model	MiniCPM-V 4.7 (1B variant)
Developer	OpenBMB
Modalities	Single image, multi-image, video
Predecessor	MiniCPM-V 4.6
Context window	Not disclosed in available material
Pricing	Not disclosed (open weights on Hugging Face)
Benchmark scores	No numerical scores in available material
Training cutoff	Not disclosed

Architecture: early visual token compression

MiniCPM-V 4.7 uses the intra-ViT early compression technique from LLaVA-UHD v4. It compresses visual tokens early inside the vision encoder, which reduces computation in the later encoder layers.

The model also supports mixed 4x/16x visual token compression. In the reference code, downsample_mode defaults to "16x", which merges tokens for efficiency. Setting it to "4x" keeps four times as many tokens for finer detail. The parameter must be passed to both apply_chat_template and generate().

Other documented processing defaults:

• max_slice_nums: default 9. OpenBMB recommends 36 for images and 1 for video.
• max_num_frames: default 128 for video.
• stack_frames: default 1. Values of 3 or 5 are recommended, adding sub-frames composited into a grid and interleaved with main frames.

Performance claims

OpenBMB says the 1B variant "advances the performance–efficiency Pareto frontier" for multimodal LLMs. It also says the model improves foundation capabilities over MiniCPM-V 4.6. These are company claims. The available text includes an overall-performance chart label for 1B Instruct but no underlying numbers, so the comparison with Qwen3.5 0.8B cannot be independently checked from this material.

Deployment and tooling

• Platforms: iOS, Android and HarmonyOS, plus open-source on-device deployment support. OpenBMB published unedited screen recordings of the 1B model running on an iPhone 17 Pro Max.
• Chips: day-0 support for Rockchip, Houmo, Intel, AMD and Arm platforms.
• Inference frameworks: SGLang, vLLM, llama.cpp, Ollama and Hugging Face Transformers.
• Fine-tuning frameworks: SWIFT and LLaMA-Factory.
• Quantization: GGUF, BNB, AWQ and GPTQ variants, with FP8 quantization support added.
• Serving: transformers serve exposes an OpenAI-compatible endpoint for quick testing and moderate-load deployment.

Running the model with Transformers requires transformers>=5.18.0. Video decoding via torchcodec can fail on some CUDA setups. For example, torch>=2.11 bundles CUDA 13.1, which conflicts with CUDA 12.x environments. OpenBMB suggests replacing it with PyAV or pinning the torch CUDA version.

Self-Replay Fine-Tuning

OpenBMB is also releasing Self-Replay Fine-Tuning. It describes the method as a low-cost way to extend model capabilities while reducing forgetting of existing ones. No evaluation data for the technique was included in the available material.

What this means

The release targets a clear gap: vision-language models small enough for phones and edge chips that still handle video and multi-image input. The selectable 4x/16x token compression is the most practical feature. Visual token count is a main cost driver for on-device multimodal inference, and a per-request switch lets developers trade detail for latency without swapping models.

The deployment coverage is broad. Day-0 support across mobile operating systems, several chip vendors, and the common quantization and serving stacks lowers integration cost for builders.

The central claim, a better efficiency frontier than Qwen3.5 0.8B, still needs verification. The available material has no benchmark numbers, latency figures, memory footprints or context length. Until those appear in the full model card or from independent evaluations, treat the performance claims as unconfirmed.