NVIDIA DGX Spark
A compact AI supercomputer introduced by NVIDIA, targeting local AI inference and development scenarios. It features high-bandwidth memory and GPU computing capabilities.
Core Facts
Timeline (last 90 days)
On the DGX Spark, the public runtime is about 20 GB unpacked and the Flash weights are about 18 GiB; Flash uses 17.75 GiB of loaded CUDA tensors, causing a peak drop of about 36.06 GiB in available host memory.
Cloudflare Clef Flash can run on an NVIDIA DGX Spark via the local /v1/systemone endpoint, providing a decision API service.
The system runs on an NVIDIA DGX Spark, equipped with Hermes Agent v0.20.6 and the Qwen 3.6-35B model.
Directly equating the DGX Spark's 64 GB configuration as the cause of Samsung's profit surge is largely an intuitive inference within the community; memory manufacturers' profit growth is driven by the entire AI server market.
In a local test on an NVIDIA DGX Spark, Embedding Gemma 2 indexed 300 animal photos and hit 165 out of 200 retrieval queries, achieving an accuracy of 82.5%.
Model training was completed on an NVIDIA DGX Spark in less than 60 minutes.
The test device is a 128G compute module (referred to in the video as the Thor T5000) claimed to have performance twice that of the NVIDIA DGX Spark.
Perplexity has launched a local computer-use agent called Portable Computer, with an exclusive first release on NVIDIA's DGX Spark device, running entirely on-device.
9B-class models run relatively smoothly on small-form-factor machines like the DGX Spark, while 31B models place higher demands on hardware.
The NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
2 more timeline events
All Facts (11)
The NVIDIA DGX Spark is equipped with the GB10 Grace Blackwell Superchip and 128 GB of unified memory.
65%UnverifiedOn the DGX Spark, the public runtime is about 20 GB unpacked and the Flash weights are about 18 GiB; Flash uses 17.75 GiB of loaded CUDA tensors, causing a peak drop of about 36.06 GiB in available host memory.
50%UnverifiedCloudflare Clef Flash can run on an NVIDIA DGX Spark via the local /v1/systemone endpoint, providing a decision API service.
50%UnverifiedThe system runs on an NVIDIA DGX Spark, equipped with Hermes Agent v0.20.6 and the Qwen 3.6-35B model.
50%UnverifiedDirectly equating the DGX Spark's 64 GB configuration as the cause of Samsung's profit surge is largely an intuitive inference within the community; memory manufacturers' profit growth is driven by the entire AI server market.
50%UnverifiedIn a local test on an NVIDIA DGX Spark, Embedding Gemma 2 indexed 300 animal photos and hit 165 out of 200 retrieval queries, achieving an accuracy of 82.5%.
50%UnverifiedModel training was completed on an NVIDIA DGX Spark in less than 60 minutes.
50%UnverifiedThe test device is a 128G compute module (referred to in the video as the Thor T5000) claimed to have performance twice that of the NVIDIA DGX Spark.
50%Unverified9B-class models run relatively smoothly on small-form-factor machines like the DGX Spark, while 31B models place higher demands on hardware.
50%UnverifiedLLM Sizer is a free tool that requires no registration to use; it compiles all Apple Silicon models starting from the M1 as well as the NVIDIA DGX Spark, includes a large number of models categorized by size, and supports custom quantization levels, context windows, budgets, and runtimes.
50%UnverifiedPerplexity has launched a local computer-use agent called Portable Computer, with an exclusive first release on NVIDIA's DGX Spark device, running entirely on-device.
50%