OpenBMB released GGUF weights for MiniCPM-V 4.7 35B-A3B, an on-device multimodal MoE model.
Hugging Face is hosting a GGUF (llama.cpp) quantization of OpenBMB's MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V vision-language series. OpenBMB says Lookahead Expert Scheduling, layer-wise quantization, and speculative sampling enable smooth on-device decoding at more than 13 tokens per second using 3 GB on an iPhone 17 Pro Max. In Instruct mode, 35B-A3B is claimed to beat Qwen3.6 35B-A3B overall while using less inference time, with early visual-token compression and mixed 4x/16x compression. Weights support SGLang, vLLM, llama.cpp, and Ollama, plus GGUF, BNB, AWQ, GPTQ, and FP8 formats and Self-Replay Fine-Tuning.
GGUF release of MiniCPM-V 4.7 35B-A3B, the series' first MoE.
Claims 13+ tokens/s in 3 GB on iPhone 17 Pro Max.
Instruct mode reportedly outscores Qwen3.6 35B-A3B in less time.
Supports iOS, Android, HarmonyOS plus Rockchip, Houmo, Intel, AMD, and Arm.
Offers GGUF, BNB, AWQ, GPTQ, and FP8, plus SRFT fine-tuning.
Full article1,905 words · extracted from huggingface.co · click to collapse
> **This repository hosts the GGUF (llama.cpp) quantized version of [MiniCPM-V 4.7 35B-A3B](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B).** For the original BF16 weights and the full model card, please refer to [openbmb/MiniCPM-V-4.7-35B-A3B](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B).
**MiniCPM-V 4.7** is the latest model in the MiniCPM-V series, available in a **35B-A3B** variant, with **35B-A3B being the first MoE model in the MiniCPM-V series**. It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include:
- 📱 **The First Open-Source MoE Model with Lossless, Smooth On-Device Inference.**
With our [**Lookahead Expert Scheduling (LES)**](https://github.com/OpenBMB/llama.cpp-omni) technique introduced alongside MiniCPM-V 4.7, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration, MiniCPM-V 4.7 35B-A3B becomes the **first open-source MoE model to run smoothly on-device without compromising performance**, achieving a decoding speed of **13+ tokens/s with just 3 GB of memory on iPhone 17 Pro Max**.
- 🔥 **A New Pareto Frontier in Performance and Efficiency.**
As model capabilities continue to advance, inference efficiency has become an important measure of a model's value. On resource-constrained devices, delivering the strongest possible capabilities within a limited resource budget calls for higher capability density. The **35B-A3B** variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier for MLLMs. In Instruct mode, 35B-A3B **achieves a higher overall score in less inference time than Qwen3.6 35B-A3B**.
- 🚀 **Efficient Architecture.**
Based on the intra-ViT early compression technique from [LLaVA-UHD v4](https://github.com/THUMAI-Lab/LLaVA-UHD-v4), MiniCPM-V 4.7 compresses visual tokens early within the vision encoder, reducing computation in subsequent encoder layers. It also supports **mixed 4x/16x visual token compression**, allowing developers to balance visual detail and inference cost across tasks.
- 💻 **Broad Platform and Chip Support.**
MiniCPM-V 4.7 supports **iOS, Android, and HarmonyOS**, together with open-source on-device deployment support. It also provides **day-0 support for chip platforms from Rockchip, Houmo, Intel, AMD, and Arm**, extending deployment options across mobile and local computing devices.
- 🛠️ **Developer Friendly.**
MiniCPM-V 4.7 supports [inference frameworks](https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/deployment_en.md#supported-inference-and-training-frameworks-and-api) including SGLang, vLLM, llama.cpp, and Ollama, along with [fine-tuning frameworks](https://github.com/OpenBMB/MiniCPM-o/blob/main/docs/deployment_en.md#supported-inference-and-training-frameworks-and-api) including SWIFT and LLaMA-Factory. It provides GGUF, BNB, AWQ, and GPTQ quantized variants and adds **FP8 quantization support**. We also release [**Self-Replay Fine-Tuning (SRFT)**](https://openbmb.cn/en/blog/self-replay-fine-tuning), a fine-tuning approach that helps developers extend model capabilities at low cost while reducing forgetting of existing capabilities, making further customization easier.
#### Smooth On-Device MoE Inference without compromising performance
An MoE model can have far more total parameters than it activates for a single token, but its on-device storage requirements are not limited to those active parameters. Different tokens can select different experts at different layers, while device memory often cannot hold all expert weights. Without a latency constraint, weights can remain on SSD or other storage, be loaded as needed, and be released after computation. This makes large-model inference possible, but repeated loading stalls can make decoding too slow for real-time interaction.
Another approach is to restrict the pool of available experts. For example, fixing an expert subset for an entire prompt can reduce loading frequency, but limits subsequent token-level routing to that subset, potentially excluding experts the model would otherwise select and reducing its capabilities. The central goal is therefore **smooth inference without compromising performance**: preserving model capability while delivering sufficient throughput under limited memory and bandwidth.
MiniCPM-V 4.7 35B-A3B addresses this challenge through three complementary techniques:
- [**Lookahead Expert Scheduling (LES)**](https://github.com/OpenBMB/llama.cpp-omni) uses cross-layer routing prediction to preload experts likely to be needed by the next layer. Among the experts selected by the actual router, it prioritizes those successfully prefetched and overlaps their computation with loading the remaining required experts. Predictions guide only prefetching and scheduling; computation still follows the model's original token-level routing, and required experts missed by the prediction are loaded and evaluated as usual.
- **Fine-grained layer-wise quantization** reduces the storage footprint and memory traffic of expert weights while preserving model capability, making expert loading better suited to on-device memory and bandwidth constraints.
- **DSpark speculative sampling integration** incorporates speculative sampling into the on-device inference process, working with expert scheduling and quantization to improve decoding efficiency.
Together, these techniques make MiniCPM-V 4.7 35B-A3B the first open-source MoE model to run smoothly on-device without compromising performance, achieving **13+ tokens/s on iPhone 17 Pro Max**.
**On-Device MoE Inference: Before and After Optimization**
<p align="center">
<a href="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/moe_speed_comparison_en.gif"><img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/moe_speed_comparison_en.gif" alt="On-device MoE inference speed comparison, with optimization on the left and without optimization on the right" width="95%"></a>
</p>
### Evaluation
**Significantly Advancing the Performance–Efficiency Pareto Frontier**
The 35B-A3B variant of MiniCPM-V 4.7 advances the performance–efficiency Pareto frontier. In Instruct mode, 35B-A3B reaches a higher overall score in less inference time than Qwen3.6 35B-A3B.
<p align="center">
<a href="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/performance_efficiency_frontier.png"><img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/performance_efficiency_frontier.png" alt="Performance versus inference time for MiniCPM-V 4.7 35B-A3B" width="95%"></a>
</p>
**Overall Performance (35B-A3B, Instruct)**
**Vision-Language**
<p align="center">
<a href="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/overall_performance_35ba3b_instruct.png"><img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/overall_performance_35ba3b_instruct.png" alt="Vision-Language performance comparison for MiniCPM-V 4.7 35B-A3B in Instruct mode" width="90%"></a>
</p>
**Language**
<p align="center">
<a href="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/overall_performance_35ba3b_language.png"><img src="https://raw.githubusercontent.com/OpenBMB/MiniCPM-o/main/assets/minicpmv4.7/overall_performance_35ba3b_language.png" alt="Language performance comparison for MiniCPM-V 4.7 35B-A3B in Instruct mode" width="90%"></a>
</p>
### Examples
MiniCPM-V 4.7 can run on phones and tablets. The clips below are unedited screen recordings captured directly on the devices.
<table align="center">
<tr>
<td align="center"><b>35B-A3B</b><br><sub>iPhone 17 Pro Max</sub></td>
> **Note on CUDA compatibility:** `torchcodec` (used for video decoding) may have compatibility issues with certain CUDA versions. For example, `torch>=2.11` bundles CUDA 13.1 by default, while environments with CUDA 12.x may encounter errors such as `RuntimeError: Could not load libtorchcodec`. Two workarounds:
>
> 1. **Replace `torchcodec` with `PyAV`** — supports both image and video inference without CUDA version constraints:
> ```bash
> pip install "transformers[torch]>=5.18.0" torchvision av
> ```
> 2. **Pin the CUDA version** when installing torch to match your environment (e.g. CUDA 12.8):