DHH called TensorFold “key to our local AI recipes for out-of-the-box agents on NVIDIA,” and the recipes are already in omarchy-local-ai. Its 0.5.0 release extends them to 64 GB machines.
What it is
TensorFold is an Apache-2.0 Zig binary that serves language models on Apple Silicon and NVIDIA, running Metal or CUDA kernels directly. Only the CUDA side is in Omarchy’s catalog for now. omarchy agent local use best walks a ranked list: on 128 GB it’s Qwen3.8 Flash Next (125B MoE, 6B active), then Nemotron 3.5 Bolt, then dense Qwen3.8 27B. On 64 GB it’s Nemotron Bolt, then Qwen3.8 27B. N1x and DGX Spark now count as one class, so memory alone decides.
The numbers
Flash Next on a 128 GB DGX Spark logged 89 tokens/s on a short prompt and 61.8 decode on a 32K prompt. These are the catalog’s own measurements, so treat them as a starting point. I don’t have N1x or Spark hardware, but this is exactly the “local agents out of the box” direction I want.
Isolate it: try new models and tools in a disposable VM or a spare machine first, with a snapshot you can roll back to.
God Bless, and May the Source Be with You!
