MiniCPM-V 4.7 vs Qwen 3.6: Who Wins Which Benchmarks?
MiniCPM-V 4.7's benchmark results against Qwen 3.6 and Gemma, showing where the Apache 2.0 vision MoE model wins, ties, and falls short.
What is MiniCPM-V 4.7 and how does it compare to Qwen 3.6?
MiniCPM-V 4.7 is an open weight vision language model released under Apache 2.0, built as a mixture of experts (MoE) with roughly 35 billion total parameters and about 3 billion active per forward pass (a 35B-A3B configuration). It’s the first MoE entry in the MiniCPM-V vision line, and it was sized and positioned to go head to head with Qwen’s similarly structured 35B-A3B model, often referred to in comparisons as Qwen 3.6. Against that direct rival, MiniCPM-V 4.7 reportedly scores higher on instruction following in instruct mode, while trading wins and losses elsewhere depending on the task category.
TL;DR
- MiniCPM-V 4.7 is an Apache 2.0 licensed vision MoE model with about 35 billion total parameters and 3 billion active, built for single image, multi-image, document, and video understanding.
- The model supports visual token compression at 4x or 6x ratios, letting developers trade visual detail for lower inference cost.
- Against the same-size Qwen 3.6 MoE model, MiniCPM-V 4.7 claims wins on instruction following and coding, plus leads on Chinese knowledge and agentic tool use.
- Gemma reportedly beats MiniCPM-V 4.7 on coding and instruction following in the broader comparison, which runs counter to the model’s own strongest claims against Qwen.
- A 27 billion parameter Qwen model reportedly leads on math and some tool calling benchmarks, showing no single model sweeps every category.
- In hands-on testing, the model handled SQL debugging and physics problems well but showed clear weaknesses in low-resource language translation and fine detail in generated visuals.
- A smaller 1 billion parameter variant also exists in the same release family, aimed at lighter deployment scenarios.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
What are the headline benchmark claims?
The core pitch behind MiniCPM-V 4.7 isn’t that it dominates across the board. According to the model’s own reported figures, it leads specifically on Chinese knowledge tasks and on agentic tool use. It stays roughly competitive on general knowledge benchmarks without pulling ahead, and it edges out its same-size Qwen 3.6 sibling on instruction following and coding.
That’s a narrower claim than “best overall,” and the data backs that framing. Gemma, a separate model family brought into the same comparison set, reportedly outperforms MiniCPM-V 4.7 on coding and instruction following, which is a notable wrinkle given that instruction following is one of the categories where MiniCPM-V 4.7 claims to beat Qwen. A 27 billion parameter Qwen variant also reportedly takes the lead on math and some tool calling tasks. The overall picture is a set of overlapping strengths rather than one model winning everything, which lines up with how most current open weight vision model releases position themselves against rivals of similar size.
How does the mixture of experts design affect performance?
MiniCPM-V 4.7’s MoE architecture activates only a fraction of its total parameters per token (about 3 billion active out of 35 billion total), which is the same basic design pattern Qwen used in its comparable release. The appeal of MoE in this context is efficiency: you get access to a larger pool of specialized parameters without paying the full inference cost of a dense 35B model on every token.
Combined with the visual token compression feature (4x or 6x), this is aimed squarely at on-device and resource constrained deployment. The model’s developers reported it running at over 13 tokens per second on an iPhone, though that figure applies to a quantized version rather than the full precision model. On more substantial hardware, such as an H100 GPU with 80GB of VRAM, the full precision model consumes around 68GB of VRAM when served through a standard transformers setup. For anyone without that kind of GPU budget, quantized versions are available that bring VRAM requirements down significantly, though at some cost to output quality.
How did MiniCPM-V 4.7 perform in hands-on testing?
Independent testing of the model (outside of its own benchmark claims) found a mixed but generally solid picture.
On a coding test asking the model to generate a single HTML file simulating a rotating Hungarian chimney cake with caramelizing sugar over glowing charcoal, the output captured the general concept reasonably well: a spiral-grooved roller, golden shading toward the heat source, and animated sparks. But the execution had rough edges. The dough rendered as a flat, blocky shape with visible pixel banding, and the glowing charcoal looked more like large blobs than a convincing coal bed.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
On a T-SQL debugging task involving a hidden bug in a query (a payment join incorrectly multiplying line amounts across multiple settled payments), the model correctly identified the bug and produced a fixed query using an EXISTS check, without disturbing the rest of the query. This lines up with the model’s own claimed strength in instruction following.
On a multilingual handwriting OCR test involving Urdu, Arabic, English, and Indonesian text with some crossed-out words, the model read English and Indonesian accurately and even named the languages with a short summary. But it skipped the Urdu line entirely, producing only a placeholder, and misread scribbled corrections. For a 35 billion parameter model being compared directly against a same-class Qwen variant, that gap stood out as a weaker-than-expected result.
On a handwritten physics problem requiring transcription of five equations and three calculation tasks, the model transcribed the equations correctly and got both final calculations right (roughly 500 nanometers for a wavelength peak and 1 eV for a photoelectron energy, including a clean derivation using E=pc). It did, however, miss identifying two of the underlying physics concepts on the sheet (Compton scattering and the relativistic Doppler equation) and rounded some intermediate figures incorrectly.
On a translation task asking the model to render a Brazilian Portuguese news headline into 80 different languages, it correctly found and transcribed the source headline and returned all 80 translations in one pass. Major languages came through with mostly sound sentence structure, though with some word-level errors even in widely spoken languages. Low-resource languages fared worse: Thai reportedly came back in Arabic script, Chinese characters leaked into Korean output, and some language pairs produced near-identical or templated results rather than genuine distinct translations.
Is MiniCPM-V 4.7 worth using over Qwen or Gemma?
The honest answer depends on what you’re building. If your priority is a permissively licensed (Apache 2.0) vision model that runs on a single GPU and handles structured reasoning tasks like SQL debugging or physics problem solving, MiniCPM-V 4.7 performs well in those lanes and matches its own benchmark claims in hands-on testing. Its instruction following held up under a real bug-fix test, which is a meaningful signal for developers building agentic or tool-use pipelines.
Where it’s weaker: fine visual detail in generated code output, low-resource language translation, and handling less common scripts in OCR tasks. If your use case leans heavily on Chinese language knowledge or agentic tool calling, the model’s own benchmarks suggest it’s a strong pick. If coding polish or broad instruction following across many task types matters more, Gemma reportedly has an edge in those specific categories, and a 27B Qwen variant leads in math.
No model in this comparison set wins everything, which is worth keeping in mind before picking one based on a single headline number.
Frequently Asked Questions
What size is MiniCPM-V 4.7 and how much VRAM does it need?
It’s a mixture of experts model with about 35 billion total parameters and 3 billion active per token (35B-A3B). The full precision version uses around 68GB of VRAM on an H100 GPU. Quantized versions are available for lower VRAM setups, including a reported 13+ tokens per second on an iPhone using a quantized build.
Does MiniCPM-V 4.7 beat Qwen 3.6 on every benchmark?
One coffee. One working app.
You bring the idea. Remy manages the project.
No. It reportedly leads on Chinese knowledge and agentic tool use, and beats the same-size Qwen 3.6 variant specifically on instruction following and coding. It doesn’t claim an overall sweep, and a separate 27B Qwen model reportedly leads on math and some tool calling tasks.
Is MiniCPM-V 4.7 good at OCR and handwriting recognition?
It handled English and Indonesian handwriting accurately in testing and correctly identified the languages, but it skipped an Urdu line entirely and misread some crossed-out corrections. It’s described by its own documentation as not primarily an OCR model, which lines up with these mixed results.
How does it perform on multilingual translation?
It successfully translated a news headline into 80 languages in a single pass, with reasonably solid results in major languages. Low-resource languages showed more frequent errors, including script mixing (such as Chinese characters appearing in Korean text) and some near-duplicate outputs between unrelated languages.
Is there a smaller version of MiniCPM-V 4.7?
Yes, the same release family includes a 1 billion parameter variant aimed at lighter-weight deployment, separate from the 35B-A3B mixture of experts model covered in most benchmark comparisons.