What if you first did the tests with MiniMax 2.7 and then followed that up by running Claude Opus 4.6 on those results. Could that provide an equal or higher quality of output at a cost in between the two you show?
It compares models of similar bench performance, where one model was trained on original data, and the other from distilation data. Clearly distilation produces inferrior results although the Minimax2.7 is much cheaper.
The aftereffect is that with the distilation model, a couple more prompts need to follow, so that the overall desired architecture (that distilation models seem to miss ) is restored. And of course the somewhat lower ability in debuging, which has its routes in the reduced ability to perceive the correct architecture of the code. This reduced perception has further negative implications, which may result in over-engineering in the future, missing or inventing bugs that do not exist etc.
The difference between the 2 models of course is not only distilation, but several other factors, the most importan in my opinion being the quality of training data and the size of the model.
What if you first did the tests with MiniMax 2.7 and then followed that up by running Claude Opus 4.6 on those results. Could that provide an equal or higher quality of output at a cost in between the two you show?
So, this compares like a BMW vs a Toyota?
It compares models of similar bench performance, where one model was trained on original data, and the other from distilation data. Clearly distilation produces inferrior results although the Minimax2.7 is much cheaper.
The aftereffect is that with the distilation model, a couple more prompts need to follow, so that the overall desired architecture (that distilation models seem to miss ) is restored. And of course the somewhat lower ability in debuging, which has its routes in the reduced ability to perceive the correct architecture of the code. This reduced perception has further negative implications, which may result in over-engineering in the future, missing or inventing bugs that do not exist etc.
The difference between the 2 models of course is not only distilation, but several other factors, the most importan in my opinion being the quality of training data and the size of the model.
How does it compare to claude sonnet 4.6? Can someone please reply? Thank you!