What does NVIDIA's GTC Berlin lineup mean for running AI models on your own hardware?

NVIDIA founder and CEO Jensen Huang opened GTC Berlin on October 21, 2026 with a keynote centered on hardware you can keep in the room. The company now sells three distinct product lines that share a common idea: the model stays on the machine.

RTX Spark puts a personal agent on a Windows laptop

RTX Spark launched on October 16. It is a laptop platform with up to 128GB of unified memory and 1 petaflop of AI compute. Reviewers from CGMag, The Verge and PCMag have already published early impressions. The platform runs personal agents. It also handles accelerated content creation and 1440p gaming at 100 frames per second.

Underdog, a desktop agent application built for fully local private AI, released its Windows version for RTX Spark the same week. Underdog uses its Husky model-specific inference engine so that user data never leaves the PC. Many RTX Spark laptops ship with an unlocked bootloader to let developers and enthusiasts customize the software stack. ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI have units available today; Acer will follow soon.

During the keynote Huang demonstrated agents on RTX Spark training Hugging Face's Microduck robot. The demo used NVIDIA Isaac Lab together with Hermes Agent and Nemotron 3.5 Bolt 74B.

DGX Spark clusters up to eight units for larger models

DGX Spark 64GB systems become available from partners on Friday, October 23. They can run Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B on a single box. When a workload needs more capacity, built-in NVIDIA ConnectX-7 networking and the NVIDIA Sync Cluster Assistant let users connect multiple units.

The October update raises the supported cluster size to eight systems in manual configuration. That yields 1TB of unified memory across the cluster, enough for frontier-class models such as GLM 5.3. Small teams can share one cluster as a backend while each developer runs concurrent sessions against the same model.

Model Launcher arrives later this month as part of NVIDIA Sync. It downloads a model and container matched to the hardware, connects the model to OpenCode, and automatically loads DGX Spark Skills. Those skills cover troubleshooting and device maintenance, plus documentation access. The same software update adds a Wi-Fi Assistant that lets users switch networks while the system runs in appliance mode.

Systems are sold directly by Acer, ASUS, Dell Technologies, GIGABYTE, HP, Lenovo, MSI and NVIDIA.

The workstation tier: 748GB coherent memory and a Windows version in December

The workstation delivers 748GB of coherent memory and up to 20 petaflops of AI performance. It can serve more than 30 users at speeds comparable to cloud models. NVIDIA Sync now supports DGX Station, simplifying connections between two Linux-based systems.

Seven new enterprise integrations let IT teams manage DGX Station through existing processes. The list includes Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, and Progress Software Chef. Tanium is also supported. These integrations accelerate approvals and reduce deployment risk while minimizing time to first token.

A Windows-based DGX Station is expected from system manufacturers in December. It will run professional applications on NVIDIA RTX PRO GPUs while large language models run on NVIDIA GB300 systems inside the same chassis. Manufacturing applications from nTop, Dassault Systèmes, Keysight Technologies, MathWorks, Rockwell Automation, Siemens, Synopsys, Unity and Altera will be supported for AI-guided design, simulation, validation and engineering automation. Agents using Hermes Agent, Perplexity or Aible can run on premises for agentic control in enterprise environments.

Three open source reference workflows bring GPT-6 Astra and NVIDIA Isaac Lab to DGX Station, along with Isaac Lab-Arena. Cosmos is also included. Developers can refine GR00T 1.7 models with reinforcement learning, train dexterous manipulation skills, and post-train Cosmos world models for synthetic data generation.

Cloud Functions enters early access on Linux-based DGX Station. Teams register local model endpoints in a shared catalog and access them through one OpenAI-compatible entry point. The model behind an endpoint can be upgraded without changing application code. Teammates use shared models without direct machine access, turning a DGX Station into a locally controlled inference service for a team.

On the show floor NVIDIA previews DSX OS running on DGX Station to simulate AI factory operations before full infrastructure exists. Another demo powers a local token factory for more than 30 software developers.

Hardware partners shipping today include ASUS, Dell Technologies, GIGABYTE, HP, MSI, Supermicro and Exxact. Windows versions are expected later this year from ASUS, Dell Technologies, GIGABYTE, HP, MSI and Supermicro.

New models and inference speedups land on day zero

Recursive Superintelligence released Recursive Flint, a 40-billion-parameter mixture-of-experts model with 2.5 billion active parameters. It is optimized for fast, accurate coding. It scores well on SWE-Bench Pro v2 and Terminal-Bench 2.1, plus SWE-bench verified. Recursive Flint was built using Recursive's Eureka Machine, which automates the full research loop and created specialized kernels that boosted generation speed on RTX Spark by 55 percent without changing model weights. Open weights are available on Hugging Face under Apache 2.0. The model runs on RTX Spark and DGX Spark systems. It also runs on GeForce RTX and RTX PRO GPUs. Developers can post-train with NeMo Automodel for supervised fine-tuning and LoRA, and with NeMo RL for group relative policy optimization.

Day-zero support arrives for NVIDIA Nemotron Bolt E4B 4B, GLM, DeepSeek, Qwen, MiniMax, Liquid AI's LFM2.5-24B-A2B and the FLUX.3 Klein diffusion model. Nemotron Bolt E4B 4GB packs natural conversation and multi-turn agentic skills into as little as 4GB of memory. It powers fast local voice assistants and interactive game characters, plus embedded interfaces on GeForce RTX GPUs, Jetson devices, RTX, and DGX Spark. It releases on Friday, October 23.

Black Forest Labs' FLUX 3 Image Klein is a 5.6-billion-parameter open weights image model for generation and editing. NVIDIA-optimized builds run on GeForce RTX GPUs with as little as 8GB of VRAM and create a 1-megapixel image in about four seconds on an RTX 5090 GPU.

MiniMax M3.1 Flash targets DGX Station and DGX Spark with long-context AI on the desktop. Liquid AI's LFM2.5-24B-A2B activates just 2 billion of its 24 billion parameters per token, making it fast enough for private tool-calling agents directly on PCs.

A new collection of quantized checkpoints optimizes RAM efficiency on local AI hardware. Each checkpoint stays close to full-precision accuracy while leaving sufficient memory for KV cache. The series starts with Qwen3.8-Flash-Next-FP4-Local on a single DGX Spark 128GB or RTX PRO 6000. GLM-5.3-Flash-FP4-Local runs on two systems. Developers can post-train on day zero with NeMo AutoModel on both DGX Spark and RTX Spark, with example recipes for Nemotron Bolt E4B and Flux 3 Image Klein, among others.

Inference engines also received updates. In llama.cpp, new speculative decoding support with MTP for Qwen 3.8 Flash Next makes end-to-end agentic tasks up to 1.4x faster on NVIDIA RTX PRO 6000 Blackwell. vLLM now includes faster speculative decoding and new FlashInfer prefill kernels for Qwen's linear-attention layers, making agentic tasks with Qwen3.8 27B up to 1.3x faster on DGX Spark. On DGX Station, GLM 5.3 Flash gets up to 1.2x higher throughput.

The lineup spans from a laptop you carry to a workstation that serves a team. What changes for an owner is the ability to pick a tier that matches the model size and the number of people who need access, then keep the data and the compute in the same room.