TensorFold recipes join Omarchy’s local AI catalog for NVIDIA N1x and DGX Spark machines
Three measured recipes now cover 64 GB and 128 GB N1x and GB10 machines, and a new Download button puts the recommended model on the agents panel’s first page.
Image: omacom/omarchy PR #18239 (GitHub)
“TensorFold is awesome! Thanks to @MiaAI_lab and @0xSero for turning me on to this project. It’s going to be key to our local AI recipes for out-of-the-box agents on NVIDIA,” DHH wrote on X on November 8, 2026. The recipes are already there: omacom/omarchy-local-ai, the repository behind the local model engine we covered when it landed, now has three of them on TensorFold, and its 0.5.0 release on November 9, 2026 extends them to 64 GB machines.
According to its README, TensorFold “serves language models from a Zig binary on Apple Silicon and NVIDIA GPUs” and runs Metal or CUDA kernels directly. The Apache-2.0 project lives under Ash Hart’s GitHub account, and he reacted on X on November 8: “Seeing that is pretty wild.” Only the CUDA side is in Omarchy’s catalog, where both TensorFold runtimes get the GPU through NVIDIA’s container device.
The catalog’s picks are in best.toml, which omarchy agent local use best walks in order. On 128 GB, the list is Qwen3.8 Flash Next, a 125B mixture-of-experts model with 6B active per token, then Nemotron 3.5 Bolt, 74B with 9B active and packed into 24 GB, then the dense Qwen3.8 27B. On 64 GB, it’s Nemotron Bolt, then Qwen3.8 27B. Pull request #5 made the N1x and the DGX Spark’s GB10 a single class, so only memory decides what runs, and a recipe measured on one qualifies on the other at the same memory.
Here’s what the catalog’s measurement records show, in tokens per second, for a short prompt, then decode and prefill on a 32K-token prompt, then a code edit:
- Flash Next, DGX Spark, 128 GB: 89.0 short, 61.8 decode, 1,690 prefill, 311.2 code edit.
- Qwen3.8 27B, DGX Spark, 128 GB: 57.7 short, 52.7 decode, 1,251 prefill, 415.4 code edit.
- Qwen3.8 27B, Dell XPS 16 N1x, 64 GB: 61.5 short, 52.9 decode, 897.6 prefill, 350.1 code edit, at a 131K context instead of 262K.
- Nemotron Bolt, Dell XPS 16 N1x, 64 GB: 90.9 short, 81.7 decode, 2,244.7 prefill, 183.8 code edit.
- Nemotron Bolt, ASUS ProArt P14 N1x, 128 GB: 86.3 short, 68.9 decode, 1,725.7 prefill, recorded on November 2, 2026 with an earlier image and no code edit test.
Those Nemotron numbers on the XPS 16 come from the record in the repository. The table in the description of PR #5 lists slightly different ones for the same laptop: 83.4 short, 82.9 decode, 2,207 prefill and 180 for the code edit.
The recipes run in rootless Podman containers on two images. The tensorfold runtime is TensorFold 1.2.1’s signed release with no patches, which the PR says beat the Mia’s AI Lab builds it tried for Flash Next and 27B on decode speed. Nemotron Bolt needs more: a fork of Mia’s AI Lab’s Bolt kit, built by Omacom. According to PR #1, the N1x’s driver copied checkpoint tensors so slowly that loading the model took 27 minutes instead of 45 seconds, until the fork uploaded them through pinned memory. That fix has been offered upstream.
“NVIDIA” has a narrow meaning here. PR #2 says the engine “detects only unified-memory NVIDIA GPUs and the image is arm64,” which leaves out RTX cards for now, and best.toml has no RTX class in 0.5.0. The same PR says the Containerfile already pins the x86_64 release for a follow-up measured on an RTX card. AMD Strix Halo machines stick with their Gufo recipes.
Finding a model is getting easier too. Omarchy #18239, merged on November 9, 2026, shows the recommended recipe on the agents panel’s first page for machines with no local model, with a Download button and a progress bar. It also installs the NVIDIA container toolkit in a floating terminal if it’s missing. The PR notes that a real download driven from the panel wasn’t checked end to end, since that needs the new engine installed.
Don’t go looking for it just yet. The 0.5.0 package was synced to omarchy-pkgs on November 9, 2026, but the local model engine didn’t make it into Omarchy 4.5.0 and the panel work was merged after that release. Both are coming in the next version, and DHH has named it: after thanking NVIDIA for a pair of DGX Sparks, he wrote on X on November 9, 2026, “Working to tune the out-of-box local AI offerings. All shipping in 4.6 shortly.”
