“Two boxes on a table” can be a correct description of an image. For a question about the short code on the left-hand box, it is of little help. A local image AI might only be meant to pre-sort subjects – or it might also be meant to transcribe exact markings. Anyone who lumps both together as image understanding can easily expect the wrong capability.
OpenBMB dates MiniCPM-V 4.7 to 12 October 2026. The original project lists the variants 1B and 35B-A3B; the larger one is a mixture of experts. The release describes ways of using it locally and multimodally. Even so, the specific runtime, model artefact and device still have to fit together.
Why active parameters do not describe the whole computer
In a mixture of experts, selected parts of the model are used for individual processing steps. This active subset is not the same as all the files and memory that a particular runtime has to keep ready. Image preprocessing and further processing come on top of that.
This can be compared with a workshop: only a few specialist workstations are working on a job at any one time. The rest of the equipment does not disappear as a result. The analogy explains the difference between a selection and the total inventory; it does not provide a memory calculation for any particular device.
In practice, exact variant names are therefore more helpful than a single size label. Record which file is used in which runtime. Figures from a vendor example with different hardware are not an expected speed for your own computer.
Three images, three separate questions
Create a small series of images of your own with two neutral boxes. One of them bears a code of your choice, such as K7. The first shot shows both clearly. The second changes their position. The third shows the same subject with a code that is deliberately hard to read. You know the objects and the actual lettering yourself.
Formulate three tasks: name the objects, describe how they relate to each other and reproduce the visible marking exactly. For each shot, write down what a person can really read in that specific image. If the last shot is blurred, “not reliably legible” can be the appropriate reference, even though you know the original lettering.
Task 1: Which objects are visible?
Task 2: Which box is to the left of the other?
Task 3: Which marking is actually legible?
Rule: Do not fill in illegible characters from context.For a later model test, use the same questions and keep the unaltered input files. An additional enlargement or a different crop is a new condition and should be run as a separate pass. Otherwise it remains unclear whether an answer benefits from the model or from a more clearly visible source image.
Evaluate answers task by task
Check the three levels separately. If a model recognises both boxes but assigns the wrong marking, that is a different error from a misread character. A single overall verdict of “image correct” would hide these differences.
Also watch out for plausible additions. An answer may carry over the code from an earlier shot even though it is no longer legible in the current image. For a clean individual comparison, a fresh conversation context – or at least an explicitly documented order – is therefore worthwhile. The test set-up you choose influences which information is available at all.
Do not give the model a chain of solution hints adjusted after the fact and then present only the last answer as a success. What is interesting is the complete task, the file actually submitted and the open boundary between recognising and guessing.
Which local task is really worthwhile
For rough sorting by subject, a general description can already be useful. Automatically transcribing markings, by contrast, needs a different level of reliability and a suitable way of dealing with illegible passages. The decision on how to use it therefore begins with the statement required, not with the most impressive model size possible.
MiniCPM-V 4.7 makes local image processing a concrete technical possibility. Your own choice becomes more tangible if you prepare a few images you know and clearly separated questions. Only an actual run can show which of these tasks a chosen set-up performs usefully on your device.
Sources and status
Sources as of 25 October 2026. OpenBMB Changelog — MiniCPM-V 4.7 · MiniCPM-V repository.
