Systems — Lab

Local AI for a desk, a studio, a department.

Single nodes and small clusters that run models on hardware you own, on a network you control, with no third-party API in the path.
Specified by memory

A local machine is sized by what fits in memory, not by a headline number.

Operations per second is the number every vendor in this category leads with, and it is the least useful one. What determines whether a machine is worth having is whether the model you intend to run holds in memory at the precision you intend to run it, and how much context you can keep alive while it does.

We specify from the workload backwards: the model families in use, the precision, the context length, the number of concurrent sessions. The configuration follows from that.

7 — 8B4 GB16 GB · 4-bit to bf1627 — 32B16 GB64 GB · 4-bit to bf1670B35 GB140 GB · 4-bit to bf16200B +100 GB480 GB · 4-bit to bf162 GB8 GB32 GB128 GB512 GBWeights resident in memory
Fig. 1 — Weights alone, before context

Weights are parameter count multiplied by bytes per parameter: roughly half a byte at 4-bit, one at 8-bit, two at bf16. So precision moves a model across a factor of four, and the key-value cache for the context window sits on top of whatever that comes to. Choosing the machine before choosing the precision is the wrong order.

Three sizes
Single node

Desk

One machine under a desk or in a studio. Runs a working model set locally, with the full toolchain installed and pinned.

Individual practitioners, researchers, engineers who want the model on the same machine as the work.

Small cluster

Studio

A handful of nodes on a shared interconnect with common storage, presenting as one pool of capacity.

Studios, research groups, engineering teams sharing capacity across a floor.

Rack-adjacent

Department

A cluster in a room with proper power and cooling, instrumented like a facility and managed like one.

Departmental IT, university groups, teams that have outgrown desks but not the building.

What comes with it
Lab — delivered configuration
Operating systemLinux or WindowsImage built to the workload, driver and accelerator stack pinned to a known-good set
RuntimeCUDA · ROCm · SYCL · Metal · Vulkan · CPUOne abstraction across the range; a model moves between Lab and Rack without porting
Model servingLocal, on first bootChat, embedding, vision, transcription, detection and forecasting runtimes installed and served
NetworkReachable without a public addressDevice-authenticated overlay; no port forwarding, no inbound exposure
TenzroJoins on first bootHardware profile detected and published to the routing layer
SupportNamed engineering contact

Tell us the models and the people. We will size it.

Design partners