This page is a short summary of the seventh solution. It is the same model, fine-tuned here with one change: a newer and larger borrowed model, Evo-1. Its fits did not learn the task, so it has no score. The full story is in the chapter on this solution.
Contents#
1. What it is#
Evo-1 is a vision-language-action model, like SmolVLA. It was published in November 2025. It has about 775 million numbers, about 70% more than SmolVLA. Solution 7 changes only the model. The plan was that the gap from solution 6 would show what the model is worth.
The next section says how it is trained.
2. How it works#
This section says what is trained and what is shared.
Only the action half of Evo-1 is trained. The picture-and-text half is frozen. The demonstrations, the training effort and the loop over a table are solution 6's and solution 5's own. Two things differ, because the model forces them. The picture is scaled from 384 to 448 pixels. And the model answers 24 numbers a waypoint, which are trimmed to 6. What it is explains each part. The next section says what this costs.
3. What it needs#
This section gives the measured costs.
It fits 122.0 million of 775.2 million numbers. It holds 3.60 GiB while answering. It needs no NVIDIA card and no change to Evo-1's code. What it needs gives the details. The next section says what it scored.
4. What it scored#
No score. It was fitted three times, with three seeds. None of the fits learned the task, so its scoring run was stopped. The results page of this chapter says why this is not a verdict on the model.