@buyergainBuilt on a sparse Mixture of Experts (MoE) architecture, it selectively activates only 11B of its 196B parameters per token. This "intelligence density" allows it to rival the reasoning depth of top-tier proprietary models, while maintaining the agility required for real-time int
i have a feeling this one's going to become my default model once the llama.cpp fixes are all ironed out. the output is great if you're willing to wait through the thinking phase and i've been able to hit about 100t/s with a 3090 + 32gb.qwen3-vl-30b-a3b-instruct will probably rem
How do we evaluate local models ourselves? I’m seeing I’m seen in consistencies with certain code outputs across difference models of the same size now according to the different benchmarks. Some of these models should be better than others which is fine but even the ones which a
Roll credits.