NVIDIA expands local AI hardware lineup with RTX Spark, DGX updates

NVIDIA used GTC Berlin to detail a broader push into local AI hardware, software and model support, with new product rollouts aimed at personal, developer and enterprise users. The announcement matters because it shows the company trying to move more AI inference and agent workflows onto local systems rather than only cloud infrastructure.

Evidence and context

In a blog post published on Wed, 21 Oct 2026, NVIDIA said RTX Spark, which launched the previous week, is intended to power personal agents on Windows PCs, while DGX Spark is aimed at developers building and scaling applications on Linux and DGX Station is positioned for enterprise teams running frontier workloads on Linux, with a Windows version expected in December. The company also said DGX Spark 64GB systems begin rolling out from partners on Fri, 23 Oct 2026, and that DGX Station systems are shipping today from ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro and Exxact.

The same announcement listed several software and workflow additions, including support for Underdog Desktop Agent on RTX Spark, NVIDIA Sync support for up to eight DGX Spark systems, and a Model Launcher update due later in October. NVIDIA also said an October software update for DGX Spark adds a Wi-Fi Assistant and updates DGX Spark Skills, while DGX Station gains seven integrations for manageability through tools including Absolute Security, Splunk by Cisco, Microsoft Intune, Perforce Puppet, Progress Software Chef and Tanium.

NVIDIA said DGX Station for Windows is expected in December from system partners and will support manufacturing applications from nTop, Dassault Systèmes, Keysight Technologies, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera. The company also said the Windows version will support agents using Hermes Agent, Perplexity or Aible on premises, though it did not give launch-day availability beyond the December timing.

Model support and performance claims

The post highlighted Recursive Flint, a new local coding model from Recursive Superintelligence, saying it is available for NVIDIA local hardware and can run at over 220 tokens per second on a 64GB RTX Spark system. NVIDIA described Flint as a 40-billion-parameter mixture-of-experts model with 2.5 billion active parameters, and said it is open-weighted on Hugging Face under an Apache 2.0 license. It also said the model’s generation speed on RTX Spark improved by 55% through specialized kernels, according to Recursive’s Eureka Machine system.

NVIDIA further said it is adding day-zero support for several models, including Nemotron Bolt E4B 4B, GLM, DeepSeek, Qwen, MiniMax, Liquid AI’s LFM2.5-24B-A2B and the FLUX.3 Klein diffusion model. The company said new checkpoints optimized for RAM efficiency are intended to help frontier-class open models run on local hardware with enough memory left for KV cache, but those claims were presented as vendor specifications rather than independent benchmarks.

For inference engines, NVIDIA said new speculative decoding support in llama.cpp can make end-to-end agentic tasks up to 1.4 times faster on NVIDIA RTX PRO 6000 Blackwell, while vLLM changes can make Qwen3.8 27B tasks up to 1.3 times faster on DGX Spark and raise GLM 5.3 Flash throughput by up to 1.2 times on DGX Station. These figures were attributed to NVIDIA and should be read as configuration-dependent performance claims.

What remains uncertain

The announcement provided product rollout dates for some systems and software updates, but it did not give pricing for the DGX products in the material supplied here. It also did not provide independent verification of the benchmark and throughput claims cited for the new models and inference engines. For RTX Spark, the supplied background notes that NVIDIA previously described starting prices from $2,399 for laptops published on 2026-10-17, but that figure applies to the RTX Spark laptop system and not to DGX products.

Beyond the dates named in the announcement, the next verified catalyst in the provided material is unknown.