MiniCPM-V 4.7 35B-A3B: Local Testing Results for OCR, SQL and Physics
Hands-on local test of MiniCPM-V 4.7, an Apache 2.0 vision MoE model, covering OCR, SQL debugging, physics reasoning and multilingual translation.
What is MiniCPM-V 4.7?
MiniCPM-V 4.7 is an open-weight vision language model from OpenBMB, released under Apache 2.0, available as a 35B-A3B mixture of experts model and a smaller 1B variant. It handles single image, multi-image, document, and video understanding, with configurable visual token compression (4x or 6x) that lets developers trade visual detail for lower inference cost. It is the first mixture of experts model in the MiniCPM-V vision line, and OpenBMB positions it as a model that can run on-device, including claims of over 13 tokens per second on an iPhone when quantized.
TL;DR
- MiniCPM-V 4.7 35B-A3B is an Apache 2.0 vision mixture of experts model from OpenBMB, built for image, multi-image, document, and video understanding.
- The model supports adjustable visual token compression (4x or 6x), letting developers choose between detail and inference cost.
- A hands-on local test on an H100 GPU used roughly 68GB of VRAM in full precision, with smaller quantized versions available for lower-memory setups.
- In coding and SQL tests, the model generated working HTML/CSS animation code and correctly diagnosed a planted bug in a T-SQL query without any agent harness.
- On handwritten multilingual OCR, it correctly read English and Indonesian text and named the languages, but skipped Urdu entirely and botched scribbled corrections.
- On a handwritten physics worksheet, it transcribed five equations correctly and solved two calculations accurately, though it missed two of the five topic areas and rounded some values incorrectly.
- On an 80-language translation task from a Brazilian Portuguese news headline, major languages read cleanly but rarer languages showed cross-contamination, such as Thai output appearing in Arabic script.
One coffee. One working app.
You bring the idea. Remy manages the project.
How does MiniCPM-V 4.7 perform on real coding tasks?
In a practical test, the model was asked to generate a single HTML file simulating a Hungarian chimney cake (kürtőskalács) rotating over glowing charcoal, with caramelizing sugar as a visual effect. The result rendered a rotating roller with a spiral groove, golden shading on the side facing the heat, and a glowing coal bed with rising sparks. The weaknesses showed up in the details: the dough appeared as a flat, blocky shape with visible pixel banding, and the “charcoal” looked more like large glowing blobs than a convincing fire bed.
This result lines up with the model’s own published benchmarks. According to OpenBMB’s reported numbers, MiniCPM-V 4.7 does not lead on every front. It performs strongly on Chinese-language knowledge tasks and agentic tool use, stays competitive on general knowledge, and beats its same-size sibling model on instruction following and coding. Notably, Google’s Gemma 4 reportedly leads on coding and instruction following in the same comparison set, and a 27B Qwen model reportedly leads on math and some tool-calling tasks. The creative coding test result, solid structure but rough visual polish, matches that mixed-but-respectable benchmark profile.
Can it find bugs in SQL queries without an agent setup?
Yes, and this was one of the stronger results in testing. The model was given a T-SQL query from SQL Server containing one deliberately planted bug, with no agentic harness or tool access, just the raw query and a request to find and fix the issue. After a reasoning pass, it correctly identified the bug: a payment join that was multiplying line amounts across orders with multiple settled payments. It then returned a corrected query using an EXISTS check, which was the expected fix, and left the rest of the query untouched. This kind of precise, surgical correction (fixing only the broken part while preserving working logic) is a meaningful signal for instruction-following quality, and OpenBMB’s own model documentation also highlights instruction following as a relative strength for this release.
How good is its OCR and handwriting recognition?
Mixed. MiniCPM-V 4.7 is not specifically branded as an OCR model, but it was tested against a handwritten sample mixing English, Indonesian, Urdu, and Arabic, including some crossed-out and scribbled words. The model accurately transcribed the English and Indonesian lines, correctly identified which languages they were written in, and produced a reasonable one-line summary. However, it completely skipped the Urdu line, substituting a placeholder of hash marks between the Arabic and Urdu sections, and misread a scribbled correction as the letter “D.”
For a 35B parameter mixture of experts model running in full precision, and one being positioned against similarly sized competitors like Qwen’s MoE vision models, the expectation going in was higher accuracy on multilingual handwriting, particularly given the model’s claimed strength in this category.
Does it hold up on scientific and physics reasoning?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Reasonably well, though not flawlessly. The model was given a handwritten sheet of physics equations and asked to transcribe them, then use them to solve three reasoning and calculation problems. It correctly transcribed five separate equations and solved two of the calculations accurately, landing on approximately 500 nanometers for a wavelength peak and 1 electron-volt for a photoelectron energy, with a clean derivation using E = pc and correctly stated physical constants.
Where it fell short: the model only identified three broad physics topic areas on the sheet, missing Compton scattering and the relativistic Doppler equation entirely, and it rounded a few figures incorrectly. Given that even larger, graduate-level models often struggle with this kind of mixed handwritten-equation-to-calculation task, the result was characterized as fair to good rather than exceptional.
How well does it handle low-resource languages?
This is where the model showed its clearest weakness. A test took a screenshot of a Brazilian Portuguese news headline and asked the model to locate the headline and translate it into 80 different languages in a single pass. It correctly found and transcribed the source headline and returned all 80 translations in order.
Quality varied sharply by language. Major world languages generally had correct sentence structure, though even some of these showed word-level mixing errors. Rarer languages fared worse: Thai output came back rendered in Arabic script, Chinese characters leaked into lines that were supposed to be Korean, and Greek lines showed similar contamination. Some translations defaulted to retaining English loanwords (such as “arbitration”) rather than translating them, and a few language pairs, like Faroese and Icelandic, produced suspiciously similar output, suggesting the model may be falling back on related languages it has more training data for.
Is MiniCPM-V 4.7 worth running locally?
For developers who need a capable, Apache 2.0 licensed vision model that can run on a single GPU, MiniCPM-V 4.7 is a reasonable option, particularly for tasks involving document understanding, code generation from visual references, or structured reasoning over mixed text-and-image inputs. It performed well above expectations on SQL debugging and physics-based reasoning, two tasks that require tight, methodical logic rather than loose creative generation.
Where it’s weaker is in exactly the areas OpenBMB’s own benchmarks suggest: fine-grained creative visual coding, and multilingual accuracy once you move past major world languages. If your use case centers on Chinese-language tasks, agentic tool use, document OCR in common languages, or technical debugging, it’s a solid fit. If you need reliable low-resource language translation or pixel-precise visual front-end generation, temper expectations.
The full-precision 35B-A3B variant required around 68GB of VRAM on an H100 GPU during testing, putting it out of reach for most consumer hardware without quantization. OpenBMB has published quantized versions for lower-memory environments, and the smaller 1B variant is aimed specifically at on-device and mobile deployment.
Frequently Asked Questions
What does “35B-A3B” mean in MiniCPM-V 4.7’s name?
It refers to the model’s mixture of experts architecture: roughly 35 billion total parameters, with about 3 billion “active” parameters used per inference pass. This is what allows larger models to run with lower compute cost than a dense model of the same total size.
How much VRAM does MiniCPM-V 4.7 need to run locally?
In full precision, the 35B-A3B variant used approximately 68GB of VRAM on an H100 GPU during testing. OpenBMB also offers quantized versions designed to run on hardware with significantly less memory, including configurations intended for on-device and mobile use.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Is MiniCPM-V 4.7 free to use commercially?
Yes. It’s released under the Apache 2.0 license, which generally permits commercial use, modification, and redistribution without the restrictions found in more limited research licenses.
How does MiniCPM-V 4.7 compare to similarly sized models like Qwen?
According to OpenBMB’s published benchmarks, MiniCPM-V 4.7 leads on Chinese-language knowledge tasks and agentic tool use, and beats its same-size sibling on instruction following and coding. It trails other models, including a Gemma 4 variant on coding and instruction following, and a 27B Qwen model on math and some tool-calling benchmarks.
What are the biggest weaknesses found in local testing?
The clearest weak points were low-resource language translation, where rarer languages showed script mixing and cross-language contamination, and precise visual/creative coding, where generated animations had correct structure but noticeable visual roughness like pixel banding and unconvincing fire effects.
