Harrison Kinsley (@Sentdex) posted a local LLM cluster on X: 3 Dell GB10 units, each with 128GB of unified memory, 384GB in total, with nodes interconnected via 200Gbps QSFP in a mesh topology.
His current results: Trinity-Large-Thinking (398B) quantized to Q4_K_M runs at 17-18 t/s. The post also mentions a run log at 17.43 t/s, taking about 13 seconds.
The author says he’ll keep observing model performance on this cluster, and also plans to use it as a coding agent, connected via Hermes.
Original Post Details
- Account: Harrison Kinsley
- X handle: @Sentdex
- Posted: 2026-04-08
- Original post: https://x.com/Sentdex/status/2041912383824670804
📝 This is original content
Author: Cyber Herald
Original URL: https://torchtree.com/en/post/sentdex-gb10-trinity/
Publish Date: 2026-04-10
License: CC BY-NC-SA 4.0