What is MiniCPM-V 4.7 (35B-A3B and 1B)?
MiniCPM-V 4.7 is the October 2026 release of OpenBMB's vision-language series, in two sizes. The 35B-A3B is the series' first mixture-of-experts model: 35.2 billion parameters, about 3 billion active per token, 256 experts with 8 routed, built on a Qwen3.5 and Qwen3.6 language backbone with a SigLIP2 vision encoder. The 1B (1.3B in OpenBMB's own table, 1.55B with its vision tower) targets phones. Both read single images, multiple images and video, and write text. Neither is related to the text-only MiniCPM 5 2B.
The vision encoder compresses image tokens early, and developers choose 4x or 16x compression per request, trading detail against speed. OpenBMB says a new runtime technique, Lookahead Expert Scheduling, lets the 35B run at more than 13 tokens a second in 3 GB of memory on an iPhone 17 Pro Max: it predicts which experts the next layer will need and loads them from storage ahead of time. That figure has not been independently measured.
Both sizes are under Apache 2.0, with quantised builds from launch. In the last 30 days the 35B drew about 325,000 Hugging Face downloads plus 275,000 for its GGUF, and the 1B GGUF about 107,000.
Is MiniCPM-V 4.7 (35B-A3B and 1B) worth it?
The 35B-A3B's strength is documents. In OpenBMB's table, at 4x compression, it scores 94.24 on OmniDocBench v1.6 against 80.17 for Qwen3.6-35B-A3B and 79.17 for Gemma 4 31B, 91.20 on OCRBench against 86.00, and 75.60 on CharXiv chart reasoning against 70.70 for Qwen3.8-27B. It also leads general tests such as MMBench (91.72 to 86.76) and MMMU (80.22 to 77.44). These are a mix of OpenBMB's runs and public reports, and we found no independent leaderboard results yet.
It loses where agents need it most. Qwen3.8-27B scores 93.31 on ScreenSpot, a test of finding interface elements on screen, to MiniCPM's 86.99, and Gemma 4 31B leads MM-IFEval instruction following at 69.59 to 65.41. Qwen3.6-35B-A3B edges it on SimpleVQA. One hands-on test by MindStudio on an H100 found clean English and Indonesian handwriting transcription but a skipped Urdu line, and scripts crossing between languages when translating into 80 of them.
The 1B is the more unusual release. At 4x compression it scores 83.17 on MMBench and 86.10 on OCRBench against 68.00 and 79.10 for Qwen3.5 0.8B, and 68.05 on ScreenSpot against 20.26 for Gemma 4 E2B, a model nearly four times its size. Gemma still wins maths-heavy visual tests such as MathVision (52.40 to 31.64), and on OmniDocBench the 1B slips slightly below its predecessor, MiniCPM-V 4.6. Its Q4_K_M language file is only 0.67 GB, but the vision file adds 1.11 GB.
The iPhone claim is the reason for the noise and the part to verify. Thirteen tokens a second in 3 GB depends on OpenBMB's fork of llama.cpp streaming experts from flash storage, and no one has published a reproduction. On ordinary hardware the 35B is a 21.17 GB Q4_K_M GGUF plus a 1.12 GB vision file, which fits a 24 GB GPU only with a modest context.
Where it earns its place
A data engineer at an insurer digitising scanned claim forms
Turn PDFs and photos of forms into structured Markdown on premises
Serve the 35B with vLLM or the 21.17 GB Q4_K_M GGUF on one 24 GB GPU, using 4x compression for dense pages. OpenBMB's OmniDocBench and OCRBench leads suit this job; spot-check handwriting and non-Latin scripts, where the MindStudio test found gaps.
An iOS developer at a travel app
Read signs, menus and tickets offline on the phone
Bundle the 1B at Q4_K_M (0.67 GB plus a 1.11 GB vision file) through OpenBMB's iOS app code. It scores 86.10 on OCRBench in OpenBMB's table; the 16x setting cuts image tokens when speed matters more than fine print.
A finance analyst team lead at an equity research firm
Summarise charts and tables in broker reports without sending them to a cloud API
Run the 35B locally under Apache 2.0, which allows commercial use with no revenue cap. Its 75.60 on CharXiv reasoning leads OpenBMB's comparison table, but check figures it extracts, since MindStudio saw rounding errors in worked answers.
Before you commit
- Setup
- pip install transformers 5.18 or later and load with AutoModelForImageTextToText, or `transformers serve` for an OpenAI-compatible endpoint; GGUF builds run in llama.cpp and Ollama with an mmproj vision file. downsample_mode must be passed to both the chat template and generate.
- Integrations
- SGLang, vLLM, llama.cpp, Ollama and OpenVINO for inference; SWIFT and LLaMA-Factory for fine-tuning; DSpark draft models for speculative decoding; day-0 support for Rockchip, Houmo, Intel, AMD and Arm chips. Tool calling works through the chat API.
- Data and privacy
- Self-hosted, so images and documents stay on your hardware. Apache 2.0 permits commercial use, modification and redistribution with no revenue cap.
- Cost at scale
- Weights are free. The 35B needs about 68 GB of VRAM at full precision (one H100 in MindStudio's test), 36.90 GB at Q8_0 or 21.17 GB at Q4_K_M; the 1B is 0.67 GB at Q4_K_M plus 1.11 GB for vision. No hosted per-token price is published.
- Limits
- Default GGUF files have thinking switched off; separate thinking builds sit in a subfolder and multiply output length. Video sampling defaults to 128 frames. The 3 GB iPhone figure relies on OpenBMB's own runtime, not stock llama.cpp.
What's good
- 94.24 on OmniDocBench v1.6, 14 points over Qwen3.6-35B-A3B (self-reported)
- 1B runs on phones and beats Qwen3.5 0.8B across the board
- Apache 2.0 for both sizes
- Selectable 4x or 16x visual token compression
What's not
- Trails Qwen3.8-27B on ScreenSpot GUI grounding
- iPhone speed claim unverified, needs OpenBMB's llama.cpp fork
- Benchmarks only as images, mixing own and public runs
- Weak on rarer languages in one independent test
Best for
Teams extracting text and tables from scanned documents and charts on their own hardware, and mobile developers who need a vision model under 2 GB.
MiniCPM-V 4.7 (35B-A3B and 1B) alternatives
MiniCPM-V 4.7 (35B-A3B and 1B): common questions
- Is MiniCPM-V 4.7 (35B-A3B and 1B) worth it?
- On OpenBMB's own numbers this is the open model to beat for reading documents and charts at the 35B size, and the 1B is the rare phone-sized vision model worth shipping. We rate it 4.0 out of 5.
- How much does MiniCPM-V 4.7 (35B-A3B and 1B) cost?
- Free Apache 2.0 weights for both sizes. No first-party hosted price published; you pay for your own hardware.
- What are the downsides of MiniCPM-V 4.7 (35B-A3B and 1B)?
- Trails Qwen3.8-27B on ScreenSpot GUI grounding; iPhone speed claim unverified, needs OpenBMB's llama.cpp fork; benchmarks only as images, mixing own and public runs; weak on rarer languages in one independent test.
- Who is MiniCPM-V 4.7 (35B-A3B and 1B) best for?
- Teams extracting text and tables from scanned documents and charts on their own hardware, and mobile developers who need a vision model under 2 GB.
- How much does MiniCPM-V 4.7 (35B-A3B and 1B) cost for a team or in production?
- Weights are free. The 35B needs about 68 GB of VRAM at full precision (one H100 in MindStudio's test), 36.90 GB at Q8_0 or 21.17 GB at Q4_K_M; the 1B is 0.67 GB at Q4_K_M plus 1.11 GB for vision. No hosted per-token price is published.
- Is MiniCPM-V 4.7 (35B-A3B and 1B) safe for company data?
- Self-hosted, so images and documents stay on your hardware. Apache 2.0 permits commercial use, modification and redistribution with no revenue cap.
- What are the best MiniCPM-V 4.7 (35B-A3B and 1B) alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4, PaddleOCR-VL-1.6, Phi-4-multimodal and MiniCPM5-2B.
Answers checked Oct 28, 2026.
Make MiniCPM-V 4.7 (35B-A3B and 1B)? Show this rating on your site
