LM Studio's 18 March guide turns NVIDIA's DGX Station GB300 from an imposing deskside computer into something closer to a privately operated model service. Its proposed route pairs the 748 GB machine with LM Studio's headless daemon and encrypted remote access, so a smaller device can use models running on the Station.
LM Studio Watch analysis
What happened
NVIDIA lists the Station with a GB300 Grace Blackwell Ultra superchip, 748 GB of coherent memory and up to 20 petaFLOPS of AI compute. LM Studio's guide installs its headless daemon, requires version 0.0.8 or newer, signs the operator in and enables LM Link. Models downloaded to the Station then become available to linked devices for loading and inference.
The worked example downloads OpenAI's gpt-oss-120b and uses it from a lightweight laptop while the Station handles inference. LM Studio also points to a same-network server route and its local interfaces for tools and projects that do not need the remote link.
Why it matters
The useful change is not merely fitting a large model into one unusually well-fed workstation. It separates the machine doing the heavy lifting from the device in someone's hands. That can make one substantial local system useful from several places while keeping the model workload on hardware the operator controls.
There is an important March-shaped brake pedal. LM Studio said LM Link was still in preview and limited to an operator's own devices; group links for colleagues were future work. The company describes the connections as end-to-end encrypted and says chats remain local, but those are first-party design claims rather than a security audit. NVIDIA's capacity figures are specifications, not proof of model speed in this setup.
Our read
A workstation with server manners is a useful idea. The clever bit is giving the furnace a remote handle, not simply admiring how much coal fits inside it. For now, though, LM Link's preview invites one person's key ring, not the whole pub table.
What to watch
- Whether group links arrive with clear roles, revocation and audit controls.
- Independent inference and network-latency tests across large open models.
- Security documentation that goes beyond the product's encryption description.
- How teams balance one powerful local node against resilience and shared demand.
Discussion spark: Would a remotely accessible workstation change your local-AI plans, or do cost, resilience and multi-user controls still favour rented infrastructure?
Sources and evidence
- Run open models on NVIDIA DGX Station GB300 (18 March 2026)
- LM Studio 0.4.6 release notes (27 February 2026)
- NVIDIA GTC 2026: Live Updates on What's Next in AI (19 March 2026)
not affiliated with or endorsed by LM Studio or Element Labs