TensorFold has a plan for running one model across a DGX Spark and a Mac, where each…
This is a dev post classified by AGK Intelligence as Hosting & infra (a launch), kept by the Dev Radar because it carries real work, not commentary.
TensorFold has a plan for running one model across a DGX Spark and a Mac, where each machine gets its own chunk of the model's layers. The usual way is tensor parallel, splits every layer across both machines, so they have to sync on every single layer. That needs a very fast link. CUDA and Metal also don't produce exactly the same numbers, so TensorFold won't mix them that way. With a layer split, one machine runs its layers and passes a small result for each token to the other, which runs the rest. For Qwen3.8 27B in bf16 that's about 20 KiB per token. Plain Ethernet handles that easily, a
Posted by netrunner (8.3k followers) 2 h ago · 24 likes · 1.2k views · view the original post on X. Kept by the Dev Radar as Hosting & infra.
More dev work like this
- How do you design web search infrastructure for tens of millions of AI agent runs a week? — @nimble_search
- Check your TrueNAS storage health — without writing API calls. — @DanKornas
- Your WordPress backup says completed. That doesn't mean everything made it. — @johnturner
- How much useful work can an AI factory get out of the power it has? — @kimmonismus
- Well Architected Framework from AWS Platform! #BigData #Analytics #DataScience #AI… — @gp_pulipaka
- Planning AI jobs on Kubernetes? Map the moving parts before adding more YAML. — @DanKornas
- RL: AWS: #DeepRacer: League PPO: The Rubber Meets The Road. #BigData #Analytics… — @gp_pulipaka
- We're reaching a point where running 100 AI agents could cost less than a single premium… — @tirthaexe
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.3k posts from 5k X accounts over the last 21 days, 9.2k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-03 17:47 UTC. Full method.