A few weeks ago, DeepSeek V4.1 Flash on two DGX Sparks did 32 tok/s. That was @MiaAI_lab's vLLM kit, and it was the best anyone had (It's still the best vLLM kit). It's now on average 2.2x times faster (more in some places). Those are some crazy performance improvements, and they are going to keep coming! This latest PR includes the following - Image input! - Several crashes and bugs fixed - Improved onboarding for first-time users @ashxhart built TensorFold, and I knew right away this would change local AI. Mia's kit was the baseline, I mostly glued it together and knocked a Spark offline once or twice. I just want to take a second and talk about how cool this all is. - Mia still has the best recipes and quants. - Ash is building a really cool community of builders that just want to make things better. - My first quant took a few tries to get right, but it thinks less, and the speed shows in DeepSeek. - 3 weeks ago I had 150 followers, and now I am contributing to the biggest projects in the local AI space. - All the people who have responded, reached out or sent DM's to work together or help each other are why this is going to succeed. - It's pushed me to develop my own open-source LocalRouter set of tools built on the same engine for people who like quick and opinionated AI tooling. Thanks, everyone. More to come soon.