DavidAU

/

Command-R-01-Ultra-NEO-V1-35B-IMATRIX-GGUF

like 13

Thank you for what you did. The effect is very interesting. The "CommandR-35B-NEO-V1-D_AU-IQ3_S-imat13.ggu" model is the peak of my PC's capabilities, I regret that I cannot test larger versions. I am able to run the standard 4_M_K version of this model (although it is larger, it takes up less memory than yours). But after several hours of testing, I can safely say that your version of "CommandR-35B-NEO-V1-D_AU-IQ3_S-imat13" is definitely much better than the standard 4_M_K version.

I see the biggest difference when using Koboltcpp - if I put certain guidelines/information in Context Data, the model remembers them very well. I think it may be even better than "Psyonic-Cetacean-Ultra-Quality-20b-GGUF-imat-plus2" - your model too (can't wait for the neo i6/i8 version).

There are surprisingly few inconsistencies when writing longer stories. I've gotten used to the fact that models often forget and change, for example, the character's hair color or other specific features. But with this model there is surprisingly little of it. Impressive!

I admire your work and enthusiasm in implementing new interesting solutions.
Thank you.