Briefs · Taipei time (GMT+8)

NVIDIA DGX Spark 64GB goes on sale October 23, with a Windows edition of DGX Station expected in December

一句重點：Clusters can grow larger too: with manual setup, as many as eight DGX Spark units can be linked for 1TB of combined memory. NVIDIA cites GLM 5.3 as an example of a frontier model in that size class that such a cluster can handle (NVIDIA blog).

發生什麼事

Running a local model inside OpenCode is something you'll soon be able to leave to Model Launcher. The tool works out whether you're on one machine or a cluster and sets up the matching model and container on its own. NVIDIA expects to fold it into NVIDIA Sync within the month (NVIDIA blog).

There's also one more model to choose from for coding locally. Recursive Flint, from Recursive Superintelligence, is a mixture-of-experts model with 40 billion parameters in total, only 2.5 billion of which are active at a time. Its weights are on Hugging Face under an Apache 2.0 license (NVIDIA blog).

01Sources

  • Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents: NVIDIA's official blog, published 2026-10-21 18:50 (Taipei time), GTC Berlin; DGX Spark 64GB on October 23, eight-unit clusters with 1TB, Model Launcher later this month, Recursive Flint specs and license, seven management integrations for DGX Station, December Windows edition with supported software and agents

More briefs →