NVIDIA RTX Spark Desktop Unveiled, Demonstrating Multiple Local AI Applications On-Site
NVIDIA showcased the RTX Spark desktop and a variety of local AI applications. The top configuration features 128GB of unified memory, a 20-core CPU, and a Blackwell GPU with 6,144 CUDA cores, with multiple PC brands rolling out products starting November 18. The live demonstrations covered remote office work using a local Agent, robotic-arm control and training, UE5 real-time rendering, Broadcast video processing, AI restoration of old videos, and other scenarios.
Why read this
AI score 62
The live demonstrations covered office work, content creation, robotics, and other scenarios, while showing how on-device operation can be applied to data retention, persistent tasks, and deployment.
Topics
Body · AI translation
Zhidongxi reported on November 18 that NVIDIA recently held an RTX Spark sharing session in Shanghai, where it explained the technical features behind its latest RTX Spark desktop products and showcased AI application cases powered by RTX Spark across fields including office work, gaming, content creation, and robotics.
Examples included locally training embodied AI models and using them to control robotic arms for grasping, natural-language robot programming and reinforcement learning training, remote work with local AI Agents, real-time rendering of massive UE5 scenes, Broadcast video processing, and AI restoration of old videos—all powered by local RTX Spark systems.
You can even try something unconventional, such as mounting an RTX Spark laptop on a robot’s head to create your own “robot companion.”
RTX Spark is already familiar to us. It is NVIDIA’s new SoC platform for local AI Agents, integrating a Blackwell GPU, a 20-core CPU jointly developed with MediaTek, and up to 128GB of unified memory. Products based on it come in laptop and desktop form factors.
The latest product showcased this time is the desktop version of RTX Spark, which will be rolled out by multiple PC brands starting November 18.
I. From asking AI questions to getting AI to do the work: RTX Spark desktop makes its debut with Day-0 support for leading new models
In NVIDIA’s view, the way people use AI is changing, shifting from “asking AI questions” to “having AI help me achieve my goals.” Users no longer need to become experts in every tool, as Agents can invoke tools on their behalf, bringing substantial changes to productivity, content creation, coding, and gaming.
Why is running AI locally so important? NVIDIA offered several reasons: personal information and data remain on the device and do not need to be uploaded; it works even without an internet connection, with all on-site demonstrations running entirely locally; and there are no cloud token, computing-power, or monthly subscription limits, leaving electricity as essentially the only cost.
It is also worth noting that the gap between local and cloud models continues to narrow. Cloud-based frontier models previously led local models by around six to eight months, but local models now lag by only about two months, making them significantly more practical.
RTX Spark is precisely the kind of local device designed to keep on-device models readily available. On the software side, it supports NVIDIA’s full suite of proprietary technologies and features, including DLSS, ray tracing, and AI acceleration.
The newly unveiled RTX Spark desktop offers up to 128GB of unified memory and likewise features a 20-core CPU and a Blackwell GPU with up to 6,144 CUDA cores.
Its main difference from the laptop version lies in its use cases. A laptop, for example, stops working when its lid is closed and is therefore better suited to short tasks. A desktop PC can operate for extended periods, making it suitable for continuously running Agents, personal assistant applications, and remote access and control, such as using a phone to directly instruct the PC to perform tasks.
Regarding the ecosystem, NVIDIA placed particular emphasis on Day-0 support for popular mainstream models. New models or applications often receive 20% of their total lifetime downloads in their first week, so they must be able to run from day one.
However, this is no easy task. Qwen alone has more than 150,000 versions on Hugging Face, and another eight new models required support during the three weeks following the end of October.
NVIDIA has partnered with mainstream Agents and applications including Copilot, Perplexity, WorkBuddy, and Qwen Office, optimizing them for RTX.
Looking at specific scenarios, more than 130 RTX-accelerated creative applications are already available for direct use, while UE5 scenes can also be rendered smoothly on RTX Spark.
For generative AI, the team combines Day-0 model support with NVFP4 quantization to make models smaller, reduce their memory footprint, and increase their speed, enabling strong performance even on laptops.
At the event, we saw several intuitive examples, such as an Agent invoking tools to complete an entire creative workflow, from a Photoshop sketch to image generation and then rendering it as a video. You do not need to master every piece of software—the Agent handles the tools itself, and you only need to “make a wish.”
Adobe applications such as Photoshop and Premiere now have native Windows on Arm versions and can benefit from RTX acceleration.
For gaming, anti-cheat systems were previously one of the biggest obstacles facing Windows on Arm. Major online game developers including the teams behind PUBG and Riot Games titles now either offer support or have pledged to do so soon. Delta Force will begin limited RTX Spark testing in a future season, and Arena Breakout will also add support.
In addition, capabilities including DLSS 5, built-in AI teammates in games, and RTX Remix can all be implemented on RTX Spark.
During the media Q&A session, NVIDIA representatives shared several real-world RTX Spark use cases. For example, in a cost-sensitive scenario, one developer built a real-time visual monitoring system. Analyzing 10 seconds of physical movement using DeepSeek-Flash costs about 10 yuan, meaning the bill for a full day of video could be “astronomical.” Moreover, if video-stream analysis is conducted in the cloud, upload and download latency could cause the system to miss several frames.
In enterprises and public institutions, automated business processes can also effectively meet requirements that data remain within the internal network. RTX Spark is also ready to use immediately in non-office settings with unstable network connections.
After the briefing, we saw a wide range of RTX AI application demonstrations in the exhibition area, where experts and project leads shared details about the underlying technologies.
In robotics, one demonstration featured a robotic arm picking up test tubes. Real-world data was collected, augmented, and used for training in a simulated environment, after which the model was deployed back to the device, enabling the robotic arm to autonomously identify the locations of test tubes and pick them up.
During the demonstration, we could converse directly in natural language with a local Agent running on RTX Spark and ask it to operate the robotic arm to complete tasks. When idle, the local Agent could also continuously perform iterative model training on its own.
At present, the available real-world scenarios and collected data are limited, so many operations must be supplemented in a simulated environment, creating a closed loop from the real world to the virtual one. This workflow has already been developed into a standalone course for teaching purposes.
Another demonstration involved controlling a robot with natural language. We could send a GitHub link directly to a local Qwen Agent running on RTX Spark. The Agent would read the technical specifications in the documentation and provide information about the robot’s joints, servos, programming language, and simulation environment.
The Agent could also automatically install a simulator and write control code based on natural-language instructions—for example, telling the robot to shake its head or kick its legs. After execution, it could report what it had done and how long it took.
In another robotics-related demonstration, an Agent on RTX Spark helped break down hundreds of pages of complex robot technical documentation and trained robots to play soccer through reinforcement learning in a virtual environment. Through reinforcement learning, two simulated robots autonomously learned to find the ball, attack, and defend in the simulation, even though they could not find the ball at all when they started.
For office productivity, one demonstration showed how a phone could remotely instruct an Agent to perform tasks. The RTX Spark desktop system can run around the clock, and users can send instructions to the Agent through WeChat on their phones, asking it to organize Feishu and WeCom messages, prepare presentation materials, and generate a link to a PowerPoint presentation.
For setup and deployment, NVIDIA has also worked extensively with teams including WorkBuddy and Qwen Office to optimize and simplify the process. The RTX section provides one-click deployment, eliminating the need for users to download models or configure the environment manually.
For creative work, the real-time editing demonstration of an extremely large UE5 scene was particularly impressive. UE5 was run on RTX Spark and loaded with a massive city scene, with ray tracing, DLSS, and other features fully enabled. Users could edit the scene and adjust its lighting in real time, and the hands-on experience was relatively smooth.
In the Broadcast AI video-processing demonstration, RTX Spark handled real-time AI noise reduction and studio audio adjustments, completely eliminating background noise. For video, it supported real-time background replacement with 4K output at 60 frames per second, as well as video upscaling and frame interpolation.
In another demonstration, a 30-year-old video in videotape format was transcribed and transformed through an AI workflow into a stylized 4K HDR video. Text-to-image models generated the characters and scenes, a ComfyUI workflow connected the various stages, Seedance generated the video, and Adobe AI handled the ending.
Conclusion: RTX Spark Becomes NVIDIA’s Key Weapon in the Race Toward the Agent Era
At the GTC Taipei conference on June 1 this year, NVIDIA officially unveiled RTX Spark, its first Arm-based PC chip. Founder Jensen Huang declared that “this is the first comprehensive overhaul of the PC product line in 40 years,” and that the agentic processing model would extend to all kinds of devices.
Now, more than five months later, not only have numerous RTX Spark products already gone on sale and entered the market, but development of its AI application ecosystem is also gradually accelerating. The many practical demonstrations showed us that local AI is rapidly becoming easier to use, more accurate, and more practical.
AI Trends