Latest AI news · Original source attached
NVIDIA cuts into DGX Spark with a 64GB model at $4,999
NVIDIA's 64GB DGX Spark runs 100B parameter models on device, pools to 128GB and 200B across two units, and ships Oct 23 from manufacturer partners.
Reported from NVIDIA
Summary written from the original first-party source — the link to it is at the end of this post.
NVIDIA added a 64GB unified memory configuration of DGX Spark as a lower cost entry point, starting at $4,999 and shipping Friday, Oct 23 through Acer, ASUS, Dell, Gigabyte, HP and MSI. It is sold exclusively through manufacturer partners rather than NVIDIA direct. The timing is explained by NVIDIA itself, which notes the 128GB DGX Spark has climbed to $6,950, roughly a 75% rise from launch price, largely on RAM price inflation.
The hardware is unchanged in kind: the GB10 Grace Blackwell Superchip, DGX OS, a ConnectX-7 NIC on every unit, and the same NVIDIA AI software stack as the 128GB model. A single unit supports models up to 100 billion parameters, and two units connect directly over a QSFP cable or a 200 GbE fabric to pool into 128GB and extend model support to 200 billion parameters, with twice the memory bandwidth.
Every performance number here is qualified. NVIDIA’s own test, using Qwen 3.8 27B, found two clustered 64GB systems delivered up to 1.7 times the performance of a single unit. There are no absolute figures published, no tokens per second, no time to first token and no latency numbers behind the 100B and 200B claims.
The software is the more interesting part for developers. NVIDIA Sync Cluster Assistant detects connected units, validates the device configuration and sets up the ConnectX-7 network automatically, so scaling from one node to two needs no software reconfiguration. NVIDIA Sync Model Launcher, due at the end of the month, will download and launch Qwen3.8 27B on a single unit or a cluster and configure OpenCode to use it. NVIDIA Agent Toolkit, CUDA-X, Nemotron open models, Ollama, vLLM and PyTorch with CUDA ship supported out of the box, with llama.cpp and LM Studio also listed.
Two caveats worth flagging. Nothing is available on Oct 2 itself, and several pieces are still unshipped, including the Model Launcher, the agent playbooks for 64GB devices and a Blender installer that NVIDIA names as one of the first major creator applications to support the platform.
(Original source: NVIDIA)
AI Slop Weekly
Get this story in context, every Sunday.
The monthly rundown of what the stories above add up to — what shipped, what changed, what to actually use. 3-minute read, free forever.
Subscribe freeNo spam. Unsubscribe anytime.