MiniCPM-V 4.7 35B-A3B

MiniCPM-V 4.7 35B-A3B is an open-weights multimodal vision-language model developed by OpenBMB, released under the permissive Apache-2.0 license. Designed specifically to push capability density on local hardware, it employs a sparse mixture-of-experts (MoE) architecture containing 35.21B total parameters while activating only 3B parameters per token. This configuration delivers the visual comprehension, reasoning, and context retention typically associated with mid-sized frontier models, but with the computation and decoding latency of a lightweight 3B network.

As a local AI model with 35.21B parameters in 2026, MiniCPM-V 4.7 35B-A3B targets a crucial gap in self-hosted artificial intelligence: high-fidelity visual understanding that fits within realistic local compute limits. It handles single-image inspection, multi-image comparative reasoning, and extensive video ingestion via a massive 262,144-token context window. Whether deployed across edge platforms, mobile hardware, Apple Silicon unified memory systems, or dedicated desktop workstations, the model prioritizes real-world throughput over brute-force compute requirements.

For engineers seeking to run MiniCPM-V 4.7 35B-A3B locally without reliance on third-party cloud APIs, the release marks a shift toward accessible multimodal inference. By combining sparse activation routines with advanced visual token compression, it challenges dense multimodal networks like Qwen2.5-VL and provides an open, highly auditable alternative to closed proprietary vision endpoints.

Understanding the MiniCPM-V 4.7 35B-A3B MoE efficiency requires separating memory footprint from runtime compute. In standard dense models, every parameter is touched on every forward pass. In this sparse MoE architecture:

• Total Parameters: 35.21B parameters, all of which must reside in storage and memory for unquantized inference.
• Active Parameters: 3B parameters per forward token pass, keeping arithmetic intensity and compute latency remarkably low.
• Context Length: 262,144 tokens, enabling the ingestion of long video clips, dense PDF archives, or multi-turn conversational histories without forced context truncation.
• Visual Encoding Engine: Built on intra-ViT early compression derived from LLaVA-UHD v4, the architecture compresses visual tokens inside the vision encoder rather than passing raw patches down the line. It features configurable 4x and 16x visual token compression ratios to balance fine-grained spatial accuracy against token generation speeds.
• Inference Optimizations: Features integrated Lookahead Expert Scheduling (LES) to minimize memory swapping latency, along with DSpark speculative decoding support for accelerated token generation.

Because only 3B parameters activate per token, generation speeds are fundamentally decoupled from the 35.21B weight scale. Once the weights are mapped into memory, matrix multiplication operations execute with the speed of a 3B model. However, developers must account for the reality that memory capacity requirements are dictated by the total parameter count, not the active count.

MiniCPM-V 4.7 35B-A3B is tuned for detailed visual reasoning, structured document processing, and temporal sequence evaluation:

• Complex Document Analysis and OCR: The intra-ViT architecture handles high-density text, architectural schematics, infographics, and intricate financial tables. The model extracts text and visual layouts cleanly, converting scanned inputs into structured formats like JSON or Markdown without missing fine annotations.
• Multi-Image Reasoning: The architecture naturally accepts multiple images within a single conversational turn. Practitioners can prompt the model to spot manufacturing defects across consecutive camera angles, compare UI revisions, or evaluate multi-page contracts.
• Extended Video Understanding: With a 262,144-token context window, the model can digest long video files converted into frame sequences. It accurately tracks temporal changes, answers targeted questions about events occurring at specific timestamps, and extracts key action summaries.
• Local Visual Agent Workflows: Low active parameter overhead makes it an ideal vision backbone for local robotic control, GUI interaction automation, or on-device screen parsing where sub-second latency is required for practical operation.

To run MiniCPM-V 4.7 35B-A3B locally, hardware planning must account for both model weights and KV cache growth over long context runs.

Memory consumption changes dramatically depending on quantization format:

• BF16 / FP16 (Raw Weights): Requires approximately 72 GB of VRAM. Needs a multi-GPU setup such as 3x or 4x RTX 4090s, an A100/H100 80GB, or an Apple Silicon Mac with 96 GB to 128 GB of unified memory.
• FP8 Quantization: Requires approximately 36 GB to 40 GB of VRAM. Fits onto a dual RTX 3090/4090 configuration (48 GB total VRAM) or an Apple Mac with 64 GB unified memory.
• 4-Bit Quantization (Q4_K_M / AWQ / GPTQ): Requires approximately 20 GB to 24 GB of VRAM for weights alone. This makes it possible to fit the base model directly onto a single 24 GB consumer GPU.

• Single Consumer GPU (RTX 3090 / RTX 4090 24GB): A single 24 GB card represents the minimum barrier for desktop deployment. Using 4-bit quantizations like AWQ or Q4_K_M, the model fits into 24 GB VRAM. However, high-resolution visual inputs and large context allocations will quickly exhaust leftover headroom. For deep context runs, a secondary GPU or a system with 32 GB+ VRAM is recommended.
• Dual GPU Workstations (2x RTX 4090 / 2x RTX 3090): The best GPU for MiniCPM-V 4.7 35B-A3B in desktop environments. 48 GB total VRAM allows you to load FP8 or Q8 builds comfortably while preserving tens of thousands of tokens of KV cache for video processing.
• Apple Silicon (M3 / M4 Max, 64GB or 128GB Unified Memory): An excellent choice for local inference. llama.cpp and native MLX implementations can run Q4 or Q8 builds with zero VRAM segmentation issues.
• On-Device / Mobile Hardware: Utilizing Lookahead Expert Scheduling and layer-wise quantization, the model can run on mobile devices such as the iPhone 17 Pro Max, holding memory usage near 3 GB through dynamic offloading while delivering 13+ tokens per second.

For general local deployment, Q4_K_M (or 4-bit AWQ) offers the best balance between visual fidelity and resource consumption. Visual encoders are sensitive to aggressive quantization, but MiniCPM-V 4.7 isolates the vision encoder at higher precision while compressing the language MoE backbone. If your hardware allows 40 GB+ of allocation, deploy FP8 to retain maximum OCR precision on low-contrast documents.

Because of its 3B active parameter profile, MiniCPM-V 4.7 35B-A3B tokens per second are significantly higher than equivalent 30B+ dense models:

• RTX 4090 (4-bit, unconstrained VRAM): Expect 35 to 55+ tokens per second during generation.
• Dual RTX 4090 (FP8): Expect 30 to 45 tokens per second.
• Apple M3/M4 Max (Q4_K_M via llama.cpp): Expect 25 to 40 tokens per second.

OpenBMB supports popular inference frameworks on day zero, including vLLM, SGLang, llama.cpp, and Ollama.

For rapid local testing, Ollama provides the simplest path:

1# Pull and run the quantized model directly via Ollama
2ollama run minicpm-v:35b-a3b

For high-throughput API endpoints or local visual agent pipelines, serve the model through vLLM or SGLang:

1vllm serve openbmb/MiniCPM-V-4.7-35B-A3B \
2  --trust-remote-code \
3  --max-model-len 32768 \
4  --gpu-memory-utilization 0.95

Both models target the 35B MoE class with low active parameter footprints. In benchmark testing, MiniCPM-V 4.7 demonstrates superior inference latency, processing document inputs faster due to its intra-ViT token compression. Qwen3.6 remains competitive in multi-step symbolic coding, but MiniCPM-V holds an edge in video context scaling and edge quantization options.

Qwen2.5-VL 32B is a dense vision-language model. While both demand roughly equivalent amounts of VRAM to store weights (around 20-22 GB in 4-bit), MiniCPM-V 4.7 generates tokens at two to three times the speed because it computes only 3B parameters per forward step instead of all 32B. Choose Qwen2.5-VL 32B if your workload requires maximum language reasoning depth across specialized academic benchmarks. Choose MiniCPM-V 4.7 35B-A3B if you need interactive speeds, video frame analysis, and smooth local decoding on consumer hardware.