With downloads of open-source AI models and frameworks expected to explode in 2026, NVIDIA announced an update to its [system/mechanism] at CES 2026.DGX Spark, a desktop AI development platformThis update not only delivers up to 2.5 times performance improvement through software optimization, but also releases new tools and workflows for agentic AI development, and can even work with the latest RTX 5090 graphics cards to accelerate the 3D creation process.
Software optimization delivers a 2.5x performance boost.
NVIDIA emphasizes that since the launch of DGX Spark, through continuous tuning with the open-source community and software stack updates, its performance has been significantly improved.
According to official data, the Qwen-235B model, through the latest TRT-LLM release and NVFP4 quantization technology, shows a performance increase of more than 2.5 times compared to its initial release. It also shows more than double the performance on Stable Diffusion 3.5 Large and PyTorch fine-tuning tasks, meaning developers can achieve faster inference and training speeds on the same hardware.
7 New Playbooks: From Dual-Machine Tuning to Nsight Copilot
To lower the development barrier, NVIDIA has released seven new Playbooks (development guides) covering all aspects from inference and fine-tuning to data science:
• Inference: Added support for VLLM, SGLang, and TRT-LLM inference, and also supports speculative decoding.
• Fine-Tuning: This is a major highlight. It now supports connecting two DGX Sparks for PyTorch fine-tuning, which is quite useful for FLUX.1 Dreambooth LoRA or LLAMA Factory fine-tuning tasks that require more display memory.
• Tools: To address the most troublesome task for developers—CUDA writing—NVIDIA will launch Nsight Copilot, an AI assistant that can run offline on DGX Spark devices. It can assist in writing CUDA core code (such as FP4 matrix multiplication) and ensure data privacy is not leaked.
Cross-industry collaboration: Accelerating MacBook video generation and RTX 5090 3D creation
DGX Spark is not only a standalone development site, but it can also serve as a powerful external accelerator.
• MacBook Pro Acceleration: Through LAN connectivity, DGX Spark can increase AI video generation speed on MacBook Pro (M4 Max and above) by 8 times. Generating 4K video using the FLUX.1 and WAN 2.2 models in ComfyUI, which previously took 8 minutes, now only takes 1 minute.
• RTX 5090 Collaboration: For mod creators, NVIDIA demonstrated a collaborative workflow combining the RTX 5090 and DGX Spark. Creators can modify mods on the RTX 5090 and let DGX Spark handle the time-consuming texture generation tasks, achieving an uninterrupted RTX Remix creation experience.
Robotics and Edge Computing: Reachy Mini Makes its Debut
In the realm of Physical AI, NVIDIA announced a collaboration with Hugging Face to power the open-source robot Reachy Mini using DGX Spark. Developers can leverage the powerful computing capabilities of DGX Spark to build AI agents and directly control this small robot, designed specifically for human-computer interaction, to conduct experiments.
In addition, NVIDIA AI Enterprise software suite will officially add support for DGX Spark at the end of January. This will unlock more application scenarios for edge computing (Edge AI), such as quality control in smart manufacturing, loss prevention detection in the retail industry, and real-time analytics in point-of-care settings.
Analysis: A Key Piece in the Puzzle of AI Development's "Ground-Based" Approach
In my opinion, as AI models become increasingly large, the costs and privacy issues of cloud-based inference are becoming more and more prominent. NVIDIA's update to DGX Spark is clearly aimed at consolidating its competitive advantage in "ground-based AI development."
By supporting dual-machine serial fine-tuning and Nsight Copilot on-device operation, NVIDIA can address the "data staying on-premises" pain point that enterprises care about most. Cross-platform collaboration with the RTX 5090 and MacBook transforms DGX Spark from a simple computing platform into an indispensable "AI computing power add-on" for creative professionals and developers. With OEM partners such as ASUS, Dell, HP, and Lenovo launching DGX Spark-designed products, we expect to see more desktop AI solutions based on this architecture by 2026.
