The news is not just another multimodal model, but the moment when an open-source MoE architecture actually reaches a phone. OpenBMB has introduced MiniCPM-V 4.7, the 35B-A3B variant, the first mixture-of-experts model in the MiniCPM-V series. The claimed figure is 13+ tokens/s on iPhone 17 Pro Max in decoding. This is not a lab number on an enterprise GPU, but a reference to a consumer device.

In MoE models, each token activates only a subset of the total parameters, which reduces per-token compute. The historical problem is that this theoretical efficiency comes at the cost of greater memory pressure and more complex scheduling, especially outside data center racks.

MiniCPM-V 4.7 tackles the problem with three levers: Lookahead Expert Scheduling, an anticipatory expert planner introduced with this model; fine-grained layer-wise quantization; and DSpark speculative sampling integration. Together, according to OpenBMB, they allow the model not to sacrifice performance compared to an uncompressed version. This should be taken as a vendor claim, but the change in approach is clear: compression is not just applied downstream; the inference pipeline is redesigned to live within the constraints of a phone.

The second signal is capability density. The team presents the model as a new Pareto frontier for MLLMs: in Instruct mode, the 35B-A3B is said to achieve a higher overall score in less inference time than Qwen3.6 35B-A3B. Beyond the individual comparison, the structural point is that the metric shifts from total parameters to usable capability per unit of time and memory.

On the architectural side, the LLaVA-UHD v4 heritage shows up in the early compression of visual tokens inside the vision encoder. MiniCPM-V 4.7 supports mixed 4x/16x visual token compression, so developers can balance visual detail and inference cost per task.

Platform support is equally relevant: iOS, Android, HarmonyOS, and, at chip level, Rockchip, Houmo, Intel, AMD and Arm with day-0 availability. This is not only a porting exercise: it means the model enters embedded and local computing ecosystems before the market consolidates around smartphones alone. For teams managing self-hosted or edge deployments, this reduces dependence on a single silicon architecture.

On the tooling side, the model is compatible with SGLang, vLLM, llama.cpp and Ollama for inference; SWIFT and LLaMA-Factory for fine-tuning. Quantized GGUF, BNB, AWQ and GPTQ variants, plus FP8 quantization support, cover much of the deployment spectrum, from laptop to edge server. FP8 is interesting because it reduces the memory footprint without pushing toward even lower-precision formats; for multimodal workloads, that can make a difference in TCO.

The less visible but perhaps most strategic addition is Self-Replay Fine-Tuning. The idea is to extend model capabilities at low cost while reducing forgetting of existing capabilities. In an on-premise context, where proprietary data stays under control, a fine-tuning path with less risk of losing general capabilities makes vertical customization more practical.

The second-order effect is a shift in incentives. Cloud multimodal inference providers lose a slice of low-latency workload that can stay on-device; embedded chip vendors gain an open-source model able to exploit their silicon without waiting for proprietary software. The third-order effect is more interesting: if capability density becomes the selection criterion, the race to total parameters loses relevance for energy- and memory-constrained deployments. The quality of quantization and scheduling infrastructure will matter more than the declared parameter count.

For those evaluating these trade-offs between visual detail, latency and memory, AI-RADAR publishes analytical frameworks at /llm-onpremise.

The immediate availability of GGUF files and support in llama.cpp/Ollama make the model testable on commodity hardware, not only on flagship devices. Historically, that type of condition has determined open-source model adoption in the field more than lab benchmarks. MiniCPM-V 4.7 does not solve data sovereignty by itself, but it moves the boundary: if a multimodal MoE can run on local silicon, sending images and video to a cloud API becomes a choice rather than a technical necessity.