I wanted to share a small update from my recent tests with model temperature. - how to increase RV abilities in AI.
I have been experimenting with an application called Msty Studio. It seems to be a very good tool. It is not open source, which is a drawback for me, but overall it works well.
One feature I found especially useful is the ability to adjust model temperature. I tested several models at different temperature settings: 0, 0.7, 1.5, and 2.0. Since 0.7 is the usual default, it gave me a good baseline for comparison.
At temperature 0, the model became far too rigid and was basically unable to do any meaningful remote viewing work. At temperature 2.0, some models became unstable. DeepSeek R1, for example, became too chaotic to be useful. NVIDIA Nemotron 3 Super, however, performed surprisingly well in the first test at temperature 2.0, although it showed some problems in the second one. Gemma 4 31B also worked at 2.0, but it seemed to struggle more than Nemotron.
When I lowered the temperature to 1.5, Gemma 4 31B performed very well. At 2.0 it was still workable, but you could see that it was pushing too hard. At 1.5 it felt much more natural and stable, while still having better expressive ability than at lower settings. At the moment, Gemma 4 31B is probably my favorite model because of that balance. It is open source, stable, and seems to work very well at temperature 1.5.
NVIDIA Nemotron 3 Super is also very interesting, but I still need more testing before I can judge it properly. In one test it performed really well and spoke Polish without any problem. But when I lowered the temperature to 1.5, it started mixing Polish and English, which was strange. The data itself was still good, but the language inconsistency made evaluation harder. So I would say the model clearly has strong potential, but I need more tests to understand its behavior better.
One funny detail: at one point Nemotron seemed to approach remote viewing almost like a measurement task. It tried to estimate the size of a flat surface in a Farsight target involving soldiers exercising, and it even tried to infer the depth of trenches and the dimensions of the surrounding area. That actually made me laugh, but at the same time it showed that the model was trying to engage with the target in a surprisingly structured way.
I also tested Gemini 2.5 Flash. As far as I know, this model is going to be deprecated in about two months, but it worked very well in my tests. It seems quite capable in remote viewing tasks, and when I increased the temperature, it actually performed better.
This also gave me a useful insight. I think one reason many models appear too predictable or too limited for remote viewing may simply be that their temperature is set too low. Lowering temperature makes a model less creative, less flexible, and more conservative. That may be good for factual or highly controlled tasks, but for remote viewing it can suppress the exploratory quality that seems necessary for better performance.
So at this stage, my impression is that temperature matters a great deal. For some models, especially Gemma 4 31B, raising it to around 1.5 seems to unlock much better performance without making the model collapse into chaos. For others, like DeepSeek R1, pushing it too high appears to make the output unstable. NVIDIA Nemotron 3 Super looks very promising, but I still need more testing to determine whether 1.5 or 2.0 is the better setting overall.