NVIDIA Floods GTC Berlin With Local AI Hardware, Models and Tooling

NVIDIA used GTC Berlin to roll out 11 local AI announcements spanning hardware, agents and models, headlined by DGX Spark 64GB systems shipping Friday, Oct. 23, and a new 40-billion-parameter local coding model, Recursive Flint, that runs at over 220 tokens per second on a 64GB RTX Spark.

The lineup covers three tiers: RTX Spark, launched Oct. 16, powers personal agents on Windows PCs; DGX Spark supports developers on Linux; and DGX Station serves enterprise teams, with a Windows version arriving in December.

What does RTX Spark change for PC users?

RTX Spark ships with up to 128GB of unified memory and 1 petaflop of AI compute, targeting personal agents, content creation and 1440p gaming at 100 frames per second. Early reviews ran positive after the October 16 release: CGMag called it "a dream machine for serious creators," The Verge said it's "designed for the future of AI," and PCMag deemed it "an extremely capable, surprisingly portable and efficient machine."

Underdog, a desktop agent app running fully on-device via its Husky model-specific inference engine, is now available for RTX Spark. Many RTX Spark laptops ship with an unlocked bootloader for developers. During the keynote, NVIDIA founder and CEO Jensen Huang demoed agents on RTX Spark training Hugging Face's Microduck robot using Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B.

ASUS, Dell, HP, Lenovo, Microsoft and MSI laptops are available now; Acer follows soon.

How much compute can DGX Spark developers get?

The new 64GB systems run models including Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Clustering changes the math: built-in ConnectX-7 networking plus NVIDIA Sync Cluster Assistant now supports eight DGX Sparks in manual configuration, delivering 1TB of unified memory — enough to run frontier-class GLM 5.3 — with concurrent sessions for small teams.

NVIDIA Sync Model Launcher arrives later this month, downloading models and containers matched to the configuration and connecting them to OpenCode. The October software update adds a Wi-Fi Assistant for appliance mode and refreshed DGX Spark Skills. Systems are sold by Acer, ASUS, Dell, GIGABYTE, HP, Lenovo, MSI and NVIDIA.

What does DGX Station add for enterprises?

DGX Station offers 748GB of coherent memory and up to 20 petaflops, enough to run GLM 5.3 and serve 30+ users at cloud-comparable speeds. Seven new integrations — Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef and Tanium — let IT manage the systems through standard enterprise tooling.

DGX Station for Windows, due in December from ASUS, Dell, GIGABYTE, HP, MSI and Supermicro, adds support for manufacturing apps from nTop, Dassault Systèmes, Keysight, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera, plus on-premises agents via Hermes Agent, Perplexity and Aible. Linux-based DGX Station systems ship today from ASUS, Dell, GIGABYTE, HP, MSI, Supermicro and Exxact.

Robotics teams get three open source reference workflows using GPT-6 Astra, Isaac Lab, Isaac Lab-Arena and Cosmos, for refining GR00T 1.7 models, training dexterity skills and post-training Cosmos world models. NVIDIA Cloud Functions reaches DGX Station in early access, turning a machine into a locally controlled inference service with a shared catalog and one OpenAI-compatible endpoint. Show-floor demos preview DSX OS simulating AI factory operations and a local token factory serving 30+ developers.

Recursive Flint: the coding headline

Recursive Superintelligence's Flint is a 40B-parameter mixture-of-experts model (2.5 billion active parameters) with top scores on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-bench verified. Its Eureka Machine research system produced kernels that boosted generation speed on RTX Spark by 55% without changing weights. Open weights are on Hugging Face under Apache 2.0, with post-training via NeMo Automodel and NeMo RL.

Day-zero models and faster inference

Day-zero support covers Nemotron Bolt E4B 4B (natural conversation and multi-turn agents in 4GB of memory, releasing Oct. 23), GLM, DeepSeek, Qwen, MiniMax, Liquid AI's LFM2.5-24B-A2B (2B of 24B parameters active per token) and Black Forest Labs' FLUX.3 Klein, a 5.6B-parameter image model producing 1-megapixel images in roughly four seconds on an RTX 5090. New FP4-quantized checkpoints start with Qwen3.8-Flash-Next-FP4-Local and GLM-5.3-Flash-FP4-Local.

Inference got faster too. llama.cpp adds MTP speculative decoding for Qwen 3.8 Flash Next, up to 1.4x on RTX PRO 6000 Blackwell; vLLM optimizations bring up to 1.3x speedups for Qwen3.8 27B on DGX Spark and 1.2x higher GLM 5.3 Flash throughput on DGX Station.