Ash Hart builds local AI software, including TensorFold, and publishes technical experiments about running models on personal hardware. His public accounts link the project and the ashhart GitHub account.[1][2][3]
Start with TensorFold →
X profile 
GitHub · ashhart Contributions
TensorFold provides local model inference with native Metal and CUDA support. Hart’s collaboration with Alex Cheema explores splitting prompt processing from token generation across a DGX Spark and a Mac: processing a prompt emphasizes computation, while generating successive tokens depends heavily on moving model data from memory. Separating those phases can let different machines do the work they suit best, with communication and configuration affecting the result.[2][3]
Tools, writing & talks
Tools
Start here
TensorFoldtensorfold.devAn open-source local inference project with Metal and CUDA execution paths.
Read the platform requirements and installation guide before comparing benchmark claims. Experience with local models helps; the documented macOS path requires Apple silicon and an appropriate model/runtime configuration.
Writing, research & talks
A collaborator’s explanation of separating prompt processing from token generation across two machines.
Collaborator explicitly named by Alex Cheema[3]Original source