MiniCPM-V 4.7 35B-A3B Tested Locally: OCR, SQL, Physics and Coding

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: MiniCPM-V 4.7 35B-A3B Tested Locally: OCR, SQL, Physics and Coding · Channel: Fahd Mirza · 09:22 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI models, Local and open-source models, Multimodal models

Summary

The video tests MiniCPM-Vision 4.7, a 35B-A3B vision model, across coding, SQL debugging, handwriting recognition, physics reasoning, and multilingual translation. It produces a convincing but imperfect HTML simulation, correctly fixes a planted SQL bug, and performs unevenly on handwritten text while solving several physics questions. It reads and transcribes a Portuguese news headline, but its translations—especially into less common languages—contain substantial errors. The reviewer concludes that the model is a strong image reader and reasoner with good SQL and physics performance, but weaker at elaborate visual coding and low-resource languages.

Key points

MiniCPM-Vision 4.7 is presented as an Apache 2-licensed 35B-A3B mixture-of-experts model, with a separate 1B variant also available.
The model supports single- and multi-image, document, and video understanding, with visual-token compression options.
The reviewer reports the model using around 68 GB of VRAM on an 80 GB H100; quantized versions are mentioned for lower-VRAM setups.
Its generated Hungarian chimney-cake simulation includes rotation, shading, glowing coals, and sparks, but the dough and fire visuals are rough.
It identifies a SQL join bug that multiplies order-line amounts and fixes it using an EXISTS check without changing the rest of the query.
In a multilingual handwriting test, it reads English and Indonesian accurately but misses the Urdu line and misreads some crossed-out writing.
It transcribes five handwritten physics equations and gets two calculations right, but misses some equations and rounds some figures incorrectly.
It extracts a headline from a Brazilian Portuguese news screenshot and returns 80 translations, though errors are common, particularly in rare languages.

Chapters

Start	Chapter	Summary
00:01	Model overview and setup	The reviewer introduces MiniCPM-Vision 4.7, its reported capabilities and performance claims, and the hardware setup used for testing.
02:05	Coding test and benchmark context	The model creates an HTML chimney-cake simulation, which has some appealing details but noticeable visual shortcomings.
04:09	SQL debugging, handwriting, and physics	The model fixes a planted SQL bug, then shows mixed results on multilingual handwriting and a stronger but imperfect performance on physics equations.
07:18	Multilingual translation and final assessment	The model translates a headline into 80 languages with many errors, leading the reviewer to characterize its strengths and weaknesses.

Quotes

developers can trade visual detail against cost. — 00:32

it has really passed this test with flying color — 04:41

It transcribed five equations correctly and got both calculations right. — 06:46

It has found and transcribed the right headline then delivered all 80 lines in order in one pass. — 08:21

strong image reader and reasoner on a single GPU, good at SQL and physics, but weaker at fancy visual code and low resource languages. — 09:22

Concepts and entities

MiniCPM-V 4.7, H100, Apache License 2.0, SQL Server, OCR, Qwen, Gemma, Fahd Mirza

Full transcript (original language)

00:01 The model which I’m downloading at the moment is mini CPM 4.7 with vision. This model is released under Apache 2. It comes in a 35 billion A3B variant which is quite popular since Quen has released a similar model a little while ago. This one is a first mixture of expert model in this whole mini CBM vision series. They also have released a 1 billion variant which I’m going to cover next.

00:32 This one handles single image multi-image document and video understanding with mixed four times or six times visual token compression which means that developers can trade visual detail against cost. The headline is quite interesting from this model. This runs smoothly on device without losing performance reaching 13 plus token per second on iPhone for example. But of

01:02 course you would need to use a quantized version. They have also reported that in the instruct mode it has even scored higher than the quen 3.6 variant in same size of 35 billion A3B. Those performances and speed figures are from them. We are going to test it out after this gets installed. You can see that I’m already serving it with transformer. Let’s proceed to the next step. And by the way, this is Fad Miraza and I welcome you to the channel. Please also

01:33 subscribe to this free weekly AI newsletter which you can subscribe at fadmmitza.substack.com. For the GPU and my machine specs, let me quickly show you. So at the moment when I run NV top you can see that it is consuming around 68 gig of VRAM on my H100 with 80 GB of VRAM. If you don’t have that much VRAM don’t worry they also have quantized versions you can install it with lesser VRAM. If you’re looking to rent a GPU or

02:05 VM on very good price you can find the link to mass compute in video description. For the first coding test, let’s go to Hungary. In this prompt, I am asking this model to create me a single HTML file to simulate a Hungarian chimney cake turning over glowing charcoal with the sugar caramelizing as it rotates. I have already sent this prompt to the model through this script which is running. So, let’s wait for it to come

02:36 back and we will see what sort of coding skills this model has got. While that happens, let me talk a bit around these benchmarks. The numbers are quite clear. This minium version is not trying to win everything. It leads the field on Chinese knowledge and on agentic tool use tasks. Stays close to the pack on general knowledge and beats its same size sibling on instruction falling and coding. Gemma for leads on coding by the way and

03:07 instruction falling, which is quite surprising. And the 27 billion coin model take math and few other stuff like tool calling. The benchmark numbers looks quite realistic. So let’s see how model performs. So the file is here. Let me open it in front of you in the browser. I’m just going to drag and drop. And there you go. We have some Hungarian food here. Looks quite good. The roller rotates with a spiral groove on the

03:39 dough. Golden shading towards the hot side which is also quite a good touch. Glowing coals and the rising sparks. Few things are also not there yet. The dough looks like a flat blocky rectangle with you know there are lot of visible pixel bands which may be okay for a non-image generation model but the real problem is around the fire. I would say because these big glowing blobs uh are

04:09 not really a charcoal bed model has tried though. Okay, so let me know what do you think about this result and if you are from Hungary really keen to know your thoughts for the next one. I am giving this model a TSQL query from SQL server with one hidden bug and ask it to find the bug and return a corrected query. No harness is involved this time and look at the result which the model has produced. After thinking for some time, it has

04:41 found the planted bug which is the payment join that multiplies lines amounts for orders with several settled payments. And then this is a fixed query with an exist check exactly as I was expecting. So it has really passed this test with flying color and it hasn’t changed the rest of the query. So instruction following is really good which they also mentioned in their model card. So really good result. Now let me do a vision test. For the vision test

05:12 I’m going to give it this handwritten image where you can see that I have various languages from bit of an udu, Arabic, English, Indonesian and uh that’s it. So and then there are some crossed out words too. So let’s see if the model is able to pick the languages and can extract these lines. It’s not an OCR model but still. Okay. So again this was the prompt and this was the image and this is the

05:43 result. I’ll also try to display the result sorry the image here for a bit. Now look at the result. The thing is that it has read the English and Indonesian handwriting accurately. Not only that, it has named those languages and gave the oneline summary as as which looks pretty good. But it totally skipped udu line and wrote only a placeholder hashed between Arabic and Uru and read the scribbled words before copy as D. So there are various mistakes

06:13 but the effort looks good. But you know what from a 35 billion parameter mixture of expert model in full precision I would expect better especially if they are comparing it with the older 3.6 in the same uh mixture of expert category. Let’s do another test. Okay, let’s do another hard one. This is another handwritten physics equation and I’m asking the model to read it. Then solve three reasoning and calculation task using what is on this sheet. Again,

06:46 this is going to test the scientific reasoning plus math capabilities. This is the prompt which I am handing over to the model and this is the result which is quite good. It transcribed five equations correctly and got both calculations right. About 500 nanometer for the VN peak and also one EV for the photoeleron with a clean E is equal to PC derivation and the constants stated. What is not right is that it named only

07:18 three broad areas and missed the Compton scattering and relativistic Doppler equations on the sheet and it rounded few figures wrongly. So that is it is quite a hard one and even larger hu fable you know grade models have uh failed to answer it fully. So I would say that it is fair to good answer if not the greatest of all time. And the final test is here. This is all about multilinguality. I’m going to give

07:50 the model a screenshot of a Brazilian Portuguese news page and ask it to find one headline and translate into 80 languages. As you can see in this prompt from English to gut and gibberish and various others from across the globe. It tests vision reading plus multilingual range especially the rarer languages. Also a big big thank you to this newspaper where I have gotten this clip and this is what I’m using. You can see their homepage on the screen. So please

08:21 do visit them. They don’t even know about this. So but thank you so much for letting me use this. Now I already have run this prompt and you can see the result. It has found and transcribed the right headline then delivered all 80 lines in order in one pass. So please check your language and let me know what do you think about the translation. So the sentence structure um looks okay in major languages but it you know there

08:51 are various mistakes there too where it has mixed up the words even in some of the major languages and the rare languages have almost all of them are broken down. For example, Thai came back in Arabic. Chinese characters leaked into the Korean and Greek lines. Though it’s a Chinese models, many index indices keep the English word arbitration and some languages are copies of their neighbors such as fairways and Icelandic seems quite similar. But anyway, let me know your

09:22 thoughts in the comments. So look, I think Min CPM uh Vision 4.7 is a strong image reader and reasoner on a single GPU, good at SQL and physics, but weaker at fancy visual code and low resource languages. Check it out. Let me know what do you think. Again, if you want to help out the channel, please become a member from the first pinned comment or from the homepage of channel. Thank you for all the