Is it a fit?Book a call

Journal3 min

A 74B model in 24 GB, a workstation with 748 GB

October brought models built to be fine-tuned and run on one GPU, and desk-side machines big enough for the largest open weights. What now fits on site.

An engineer crouched at an open server rack, checking a cable

For a model to stay on site, it has to fit. Memory is what decides the size of model a company can run on its own premises, and in October it grew on both sides: smaller models that do more, and machines that hold more.

Models made to fit

NVIDIA released Nemotron 3.5 Bolt, a 74 billion parameter model with 9 billion active, in a version quantised to fit in 24 GB of memory, the size of a single workstation GPU. Its weights are meant mainly for further training on the user’s own data, under a licence cleared for commercial use. A small sibling, Bolt E4B, is a 6 GB file for voice and device assistants.

Machines that hold more

  • DGX Station: 748 GB of memory, shipping now from seven manufacturers.
  • DGX Spark: the 64 GB version ships from 23 October, and up to eight can be joined for 1 TB.
  • Mac Studio with M5 Ultra: up to 512 GB of unified memory, the largest configuration announced for late October.

What it means for Eigna Local

Not long ago, keeping a capable model on site meant a data centre project. Now a company can choose: a mid-size expert fine-tuned on its own work, on one GPU it already owns, or one of the large open models on a desk-side machine.

That is the class of hardware Eigna Local runs on: your own servers today. The Eigna Box, our own home for the model, is planned, and the first one is built by us. We will publish what each machine sustains on real tasks before we quote a throughput.

Notes

  1. NVIDIA, Nemotron 3.5 Bolt 74B-A9B, 15 October 2026, and Bolt E4B, 23 October 2026, both under OpenMDW-1.1.
  2. NVIDIA, GTC Berlin, 21 October 2026.
  3. Apple Newsroom, 22 September 2026.

Read next

Newsletter

Release notes.

Once a month: what we released, what we held, and why. Short, dated and true, losses included.

Double opt-in. One click to leave. We never share the list.

Each issue holds

  1. One release or testWhat passed the gate, what was held, and the reason in one line.
  2. One Journal pieceAn explainer from the people who build it.
  3. One numberWith its baseline and its date, never on its own.

The first issue comes in November 2026.

Start

Bring one task. See the proof first.

In about four weeks, on your own data in a sealed setup, we show whether an Eigna expert beats your baseline. If it does not, you keep the report.

Book a callRequest a Test Lab