NVIDIA Expands Local AI With RTX Spark, DGX Clusters and Recursive Flint
NVIDIA expands local AI with RTX Spark, eight-unit DGX clusters, Windows DGX Station, Recursive Flint and faster models at GTC Berlin for developers.
Summary
At GTC Berlin on October 21, 2026, NVIDIA expanded local AI. RTX Spark, released October 16, delivers up to 128GB unified memory, 1 petaflop of AI compute and 1440p gaming at 100 fps. Underdog’s Husky-powered agent now keeps models and data on-device, while many laptops have unlocked bootloaders. Jensen Huang trained Hugging Face’s Microduck robot using Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B.
DGX Spark 64GB arrives from partners Friday, October 23, supporting Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Manual eight-unit ConnectX-7 and Sync clusters provide 1TB memory for larger models, longer contexts and concurrent agents. October’s Sync Model Launcher downloads suitable models and containers, connects OpenCode and loads DGX Spark Skills; the update also adds Wi-Fi Assistant. DGX Station offers 748GB, 20 petaflops and cloud-comparable service for 30+ users, while Sync connects two Linux units. Seven integrations from Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef and Tanium add enterprise control.
Windows DGX Station arrives in December with apps from nTop, Dassault Systèmes, Keysight Technologies, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera, plus on-premises Hermes Agent, Perplexity and Aible. Three open workflows combine GPT-6 Astra, Isaac Lab, Isaac Lab-Arena, Cosmos and GR00T 1.7; Cloud Functions enters Linux early access. Recursive Superintelligence’s Apache 2.0 Flint, a 40-billion-parameter mixture-of-experts model with 2.5 billion active parameters, exceeds 220 tokens per second on 64GB RTX Spark; Eureka Machine kernels raised speed 55%. Day-zero support spans Nemotron Bolt E4B 4GB, GLM, DeepSeek, Qwen, MiniMax, LFM2.5-24B-A2B and FLUX.3 Klein; Bolt launches October 23, while 5.6-billion-parameter FLUX creates a 1-megapixel image in about four seconds on RTX 5090 with 8GB VRAM. llama.cpp boosts Qwen tasks 1.4x; vLLM raises Qwen3.8 27B 1.3x and GLM 5.3 Flash throughput 1.2x.
Positives
- RTX Spark combines up to 128GB unified memory and 1 petaflop of AI compute with fully local Underdog agents that keep user data on-device.
- Eight DGX Spark units provide 1TB unified memory for larger models, longer contexts, concurrent agents and shared development workloads.
- DGX Station’s 748GB memory and 20 petaflops can serve more than 30 users at speeds comparable to cloud models.
- Recursive Flint exceeds 220 tokens per second on 64GB RTX Spark, with automated Eureka Machine kernels improving generation speed by 55%.
- Apache 2.0 open weights let developers run Recursive Flint on RTX Spark, DGX Spark, GeForce RTX and RTX PRO GPUs.
- ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro and Exxact are already shipping DGX Station systems.
Risks & concerns
- Eight-unit DGX Spark clusters still require manual configuration despite their 1TB memory capacity.
- DGX Station for Windows will not arrive from manufacturing partners until December 2026.
- NVIDIA Cloud Functions is initially limited to early access on Linux-based DGX Station systems.
- Running FLUX.3 Klein locally requires at least 8GB VRAM, while larger optimized checkpoints can require two systems.