NVIDIA Expands Local AI Hardware at GTC Berlin

At GTC Berlin, NVIDIA is showcasing local AI hardware and software announcements across RTX Spark, DGX Spark and DGX Station. RTX Spark runs personal agents on Windows PCs; DGX Spark supports developers building and scaling applications on Linux; DGX Station serves enterprise teams running frontier workloads on Linux, with a Windows version expected in December. NVIDIA also announced support for eight DGX Sparks, seven new DGX Station integrations, three open source robotics workflows and day-zero support for several AI models.

NVIDIA says RTX Spark, released on October 16, offers up to 128GB of unified memory and 1 petaflop of AI compute, alongside content creation and 1440p gaming at 100 frames per second. The Underdog desktop agent is now available for RTX Spark and runs models entirely on the device through its Husky inference engine, keeping user data on the PC. NVIDIA says many RTX Spark laptops have an unlocked bootloader for developers and enthusiasts to customize local agents. During the GTC keynote, Jensen Huang demonstrated agents on RTX Spark training Hugging Face’s Microduck robot with NVIDIA Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B. Laptops from ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI are available today, with Acer expected soon.

DGX Spark systems with 64GB of memory start rolling out from partners on Friday, October 23. NVIDIA says supported models include Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Developers can connect up to eight DGX Spark units using ConnectX-7 networking and NVIDIA Sync, providing 1TB of unified memory for larger models, longer contexts and concurrent agents. NVIDIA says this setup can run models such as GLM 5.3 and support small teams sharing a cluster. Model Launcher, due later this month, will download a model and matching container for a single system or cluster, then connect it to OpenCode. An October software update adds a Wi-Fi Assistant for switching networks in appliance mode and workflow guidance through DGX Spark Skills.

DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance; NVIDIA says it can run GLM 5.3 and serve more than 30 users at speeds comparable to cloud models. NVIDIA Sync now connects two Linux-based systems, while seven integrations let IT teams manage stations through enterprise setup, management and compliance processes. DGX Station for Windows is expected from partners in December, bringing local models and agents into Windows software and manufacturing applications from companies including Siemens, MathWorks and Dassault Systèmes. The Windows system is also planned to support on-premises agents from Hermes Agent, Perplexity and Aible.

NVIDIA released three open source workflows for robotics teams to refine GR00T 1.7 with reinforcement learning, train dexterous manipulation skills and post-train Cosmos world models for synthetic data generation. Cloud Functions is coming in early access to Linux DGX Station, where teams can register local model endpoints in a shared catalog and access them through one OpenAI-compatible entry point. NVIDIA also previewed DSX OS simulations of AI factory operations and a local token factory serving more than 30 software developers.

Recursive Superintelligence released Recursive Flint, a local coding model with 40 billion parameters, of which 2.5 billion are active. The company reports inference above 220 tokens per second on a 64GB RTX Spark and says specialized kernels raised generation speed on the device by 55% without changing model weights. The open weights are available on Hugging Face under the Apache 2.0 license. Developers can run Flint on RTX Spark, DGX Spark, GeForce RTX and RTX PRO GPUs, and post-train it using NeMo tools.

NVIDIA announced day-zero support for Nemotron Bolt E4B, GLM, DeepSeek, Qwen, MiniMax, Liquid AI’s LFM2.5-24B-A2B and FLUX.3 Klein. Bolt E4B fits in as little as 4GB of memory and is scheduled for release on October 23. FLUX.3 Klein has 5.6 billion parameters and can generate a 1-megapixel image in about four seconds on an RTX 5090, using as little as 8GB of VRAM. NVIDIA’s optimized local checkpoints use quantization to reduce memory needs while staying close to full-precision accuracy.

NVIDIA reported inference speedups from work with the llama.cpp and vLLM communities: speculative decoding with MTP makes agentic tasks using Qwen 3.8 Flash Next up to 1.4x faster on RTX PRO

At GTC Berlin, NVIDIA is showcasing local AI hardware and software announcements across RTX Spark, DGX Spark and DGX Station. RTX Spark runs personal agents on Windows PCs; DGX Spark supports developers building and scaling applications on Linux; DGX Station serves enterprise teams running frontier workloads on Linux, with a Windows version expected in December. NVIDIA also announced support for eight DGX Sparks, seven new DGX Station integrations, three open source robotics workflows and day-zero support for several AI models.

NVIDIA says RTX Spark, released on October 16, offers up to 128GB of unified memory and 1 petaflop of AI compute, alongside content creation and 1440p gaming at 100 frames per second. The Underdog desktop agent is now available for RTX Spark and runs models entirely on the device through its Husky inference engine, keeping user data on the PC. NVIDIA says many RTX Spark laptops have an unlocked bootloader for developers and enthusiasts to customize local agents. During the GTC keynote, Jensen Huang demonstrated agents on RTX Spark training Hugging Face’s Microduck robot with NVIDIA Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B. Laptops from ASUS, Dell Technologies, HP, Lenovo, Microsoft and MSI are available today, with Acer expected soon.

DGX Spark systems with 64GB of memory start rolling out from partners on Friday, October 23. NVIDIA says supported models include Nemotron 3.5 Bolt 74B 24GB, Qwen3.8 27B and Gemma 4 26B-A4B. Developers can connect up to eight DGX Spark units using ConnectX-7 networking and NVIDIA Sync, providing 1TB of unified memory for larger models, longer contexts and concurrent agents. NVIDIA says this setup can run models such as GLM 5.3 and support small teams sharing a cluster. Model Launcher, due later this month, will download a model and matching container for a single system or cluster, then connect it to OpenCode. An October software update adds a Wi-Fi Assistant for switching networks in appliance mode and workflow guidance through DGX Spark Skills.

DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance; NVIDIA says it can run GLM 5.3 and serve more than 30 users at speeds comparable to cloud models. NVIDIA Sync now connects two Linux-based systems, while seven integrations let IT teams manage stations through enterprise setup, management and compliance processes. DGX Station for Windows is expected from partners in December, bringing local models and agents into Windows software and manufacturing applications from companies including Siemens, MathWorks and Dassault Systèmes. The Windows system is also planned to support on-premises agents from Hermes Agent, Perplexity and Aible.

NVIDIA released three open source workflows for robotics teams to refine GR00T 1.7 with reinforcement learning, train dexterous manipulation skills and post-train Cosmos world models for synthetic data generation. Cloud Functions is coming in early access to Linux DGX Station, where teams can register local model endpoints in a shared catalog and access them through one OpenAI-compatible entry point. NVIDIA also previewed DSX OS simulations of AI factory operations and a local token factory serving more than 30 software developers.

Recursive Superintelligence released Recursive Flint, a local coding model with 40 billion parameters, of which 2.5 billion are active. The company reports inference above 220 tokens per second on a 64GB RTX Spark and says specialized kernels raised generation speed on the device by 55% without changing model weights. The open weights are available on Hugging Face under the Apache 2.0 license. Developers can run Flint on RTX Spark, DGX Spark, GeForce RTX and RTX PRO GPUs, and post-train it using NeMo tools.

NVIDIA announced day-zero support for Nemotron Bolt E4B, GLM, DeepSeek, Qwen, MiniMax, Liquid AI’s LFM2.5-24B-A2B and FLUX.3 Klein. Bolt E4B fits in as little as 4GB of memory and is scheduled for release on October 23. FLUX.3 Klein has 5.6 billion parameters and can generate a 1-megapixel image in about four seconds on an RTX 5090, using as little as 8GB of VRAM. NVIDIA’s optimized local checkpoints use quantization to reduce memory needs while staying close to full-precision accuracy.

NVIDIA reported inference speedups from work with the llama.cpp and vLLM communities: speculative decoding with MTP makes agentic tasks using Qwen 3.8 Flash Next up to 1.4x faster on RTX PRO