Benchmarks of my models
I try to benchmark all my models prior to making them public. This means I often have to fix real problems with vendor chat templates or llama.cpp compatibility before I release the models themselves. I provide working commands to run these generally in llama.cpp because that's what I run. My models are, in general, sized to fit on either a 16GB GPU or at 128GB GB10 (which used to be moderately affordable). This is the hardware I've got. I'm happy to target another size if anybody wants. I also play with even smaller coding agents just out of curiosity and I've lately been amazed at what even 2B and 4B models can do in actual agentic coding loops.
I contrast this to the excellent work others do providing sweeps of sizes for a given model, eg q2-q8. Others I follow such as @unsloth and @bartowski are some of my favorites, and their models sometimes exceed the performance of equally sized vendor models in my experience. Please check out my work and comment or like if you find something useful. I don't have sponsors or ads, but it makes me feel good.