MiniCPM-V 4.7: Vision Models, Phone Memory and Running Costs

The 1B model is a compact starting point for an image-and-text task. The larger 35B-A3B model activates roughly 3B parameters per token, while the full expert collection remains part of storage and loading. Both have Apache 2.0 model cards.

Open the matching model card, not a similarly named MiniCPM-o or older MiniCPM-V package. Record the checkpoint revision, quantization and vision components alongside your prompt. A text-only load is not a successful multimodal setup.

Start with the official hosted demo and one image you are allowed to upload. Ask a bounded question with a checkable answer, such as a number and its label. Compare the answer with the source image before trying a larger document or video.

For local use, follow the current Transformers or supported-runner instructions in the upstream guide for the exact variant. Keep the text and vision components together. Use its image example before adapting message construction, thinking settings or frame sampling.

The examples expose 4× or 16× visual-token compression. Higher compression reduces the visual representation; test whether small text and chart detail survive on your actual inputs. Keep compression and thinking settings fixed when comparing speed or answer quality.

OpenBMB reports about 3GB memory and 13+ decoding tokens/s on iPhone 17 Pro Max for its optimized 35B-A3B path with Lookahead Expert Scheduling. The project also provides DSpark speculative decoding support. LES overlaps expert loading and computation rather than keeping an ordinary full-precision checkpoint entirely in that memory.

This is a vendor-reported configuration, not a minimum RAM figure for every runtime. Storage, quantized expert layout, visual working memory and the runner still matter. A desktop Transformers load and an LES phone deployment are different memory plans.

The litert-community converter card reports a Galaxy S26 test of its MiniCPM-V-4.7-1B-VL_int8.litertlm bundle. Its text run uses 1,024 prompt tokens and 256 decoding tokens over three iterations. The image task is a 448 × 448 photo with an 88-token prompt, measured separately.

These are the converter maintainer’s measurements, not OpenBMB’s 35B-A3B LES result. Text TTFT, image first-token time, decoding and initialization are separate counters. Keep the artifact, device and execution backend with the number.

The converted bundle supports one image per conversation at 16× compression, with thinking off; video, multi-image input and the finer 4× mode need another runtime. Its Galaxy GPU initialization takes 50.8 seconds before the reported 0.93-second image response interval, so include cold loading when planning short sessions.

Running downloaded weights avoids a provider’s per-request API charge. Equipment, electricity, storage and elapsed time remain costs. For illustration, an assumed 500W whole-system draw running for two hours at $0.20/kWh uses 1kWh and costs $0.20 in electricity; those are assumed inputs, not a MiniCPM benchmark.

Use the existing-hardware electricity option with your measured whole-machine watts and elapsed hours. Include loading, visual processing and retries. The published phone numbers do not establish desktop throughput or a GPU rental cost per document.