Can a 64 GB DGX Spark run Qwen3.8-27B?

Yes, in NVFP4 without the 1M window: two requests at a time, 59,625 tokens of KV cache, so one prompt of up to 54,855 tokens, at 67.2 tok/s (greedy median). Load it from the cockpit.

What is the best model for a 64 GB DGX Spark?

For most work, NVIDIA's Nemotron 3.5 Bolt: it reads images and video, answers at 110.8 tok/s (greedy median) and keeps 667,874 tokens of KV cache, 2.55 times its 262,144-token window.

Can a 64 GB DGX Spark make images or video?

Not with this installer. The image and video lanes need a 128 GB DGX Spark, and the installer refuses them on a 64 GB box before downloading anything.

Is 64 GB enough for Claude Code or opencode?

Yes. Bolt's pool holds the largest agent step measured on the project (43,863 tokens) many times over. On the 27B, a step that large can be refused as too long, and Claude Code and opencode compact and carry on.

What to run on a 64 GB DGX Spark: measured models

Serve NVIDIA's Nemotron 3.5 Bolt: it reads images and video and answers at 110.8 tok/s (greedy median), with 667,874 tokens of KV cache, 2.55 times its 262,144-token window. Keep Qwen3.8-27B one Load away for text and code: 67.2 tok/s, two requests at a time. One command installs exactly that, and refuses what needs 128 GB before a byte is downloaded.

What fits, measured

Lane	On a 64 GB DGX Spark
Nemotron 3.5 Bolt	The lane it serves, with a memory share of its own: 667,874 tokens of KV (2.55 times its 262,144-token window), 110.8 tok/s greedy median on `./bench.sh`.
Qwen3.8-27B in NVFP4	A Load away, in a unit sized for 64 GB: two requests at a time, 59,625 tokens of KV (one prompt of up to 54,855 tokens), 67.2 tok/s greedy median.
Speech	As on any box, on the CPU: Qwen3-ASR hears, Kokoro speaks.
Needs 128 GB	The flash lane, the 1M window of the 27B, the FP8 targets, images, video, Kolibri-1, Nemotron 3.5 Super VL, Flash FP4-Local and Ornith 1.5 122B are refused, as they need a 128 GB DGX Spark.

Measured on: an ASUS Ascent GX10 with 128 GB, at the memory a 64 GB box gives each engine, on 2026-10-24
How: every setting tried is in the changelog of v1.40.0
Source: README, On a 64 GB DGX Spark and CHANGELOG

How the installer knows the box has 64 GB

It reads the box's class from its memory: a 128 GB box from 110 GiB, a 64 GB DGX Spark from 48 GiB, and below that it stops at its first step and says how many GiB the box has. Each refusal comes before a byte is fetched: in the installer, in each lane's installer, in the model switch, and so in the cockpit's Load, where a lane a 64 GB box cannot hold says it needs a 128 GB DGX Spark. The cockpit names the box from its firmware, for example "ASUS Ascent GX10, 64 GB".

Coding agents on 64 GB

Bolt is the better fit for long agent sessions on 64 GB: its pool holds, many times over, the largest of 2,156 opencode steps measured on the project, 43,863 tokens. The 27B's pool holds an agent's turns but can refuse a step that large as too long, and Claude Code and opencode then compact and carry on. The installer fits opencode's limits to the pool the 27B boots to.

Which machines come with 64 GB

NVIDIA's 64 GB configuration has the same GB10 chip and the same 273 GB/s of memory bandwidth as the 128 GB one, and NVIDIA rates it for models of up to 100B parameters (NVIDIA's DGX Spark page). It is sold only by NVIDIA's partners: from October 23, 2026, by Acer, ASUS, Dell, GIGABYTE, HP and MSI (NVIDIA's announcement). See every GB10 machine.

Two 64 GB boxes

NVIDIA rates two linked 64 GB units for models of up to 200B parameters. This project serves one box today. A two-box mode is being designed with an owner of a linked pair, in discussion #45.

Install it

On a 64 GB box, the first install ends on Bolt. The same command updates it later.