RTX Spark laptops, available from October 16 starting at $2,399, are receiving positive early reviews. CGMag called it 'a dream machine for serious creators,' The Verge described it as 'designed for the future of AI,' and PCMag called it 'an extremely capable, surprisingly portable and efficient machine'. A new desktop agent app, Underdog, runs models entirely on-device using its Husky inference engine, keeping user data local. Jensen Huang demonstrated using RTX Spark AI agents to train a Hugging Face Microduck robot via NVIDIA Isaac Lab. Many RTX Spark laptops ship with an unlocked bootloader to support developer customization. Acer is listed as coming soon, with current availability from ASUS, Dell, HP, Lenovo, Microsoft, and MSI.
The RTX Spark chip combines a Blackwell RTX GPU with up to 6,144 cores and a Grace CPU at 600 GB/s, delivering one petaflop of FP4 AI performance with up to 128GB of LPDDR5x unified memory. Microsoft's Surface Laptop Ultra starts at $2,599 and the Surface RTX Spark Dev Box at $5,999. Launch celebrations were held in Berlin, Beijing, Istanbul, Seoul, San Francisco, and Taipei; Jensen Huang made a surprise appearance at the Berlin event, distributing popcorn, taking selfies, and giving away five personally signed laptops; Berlin's Festival of Lights featured RTX Spark-themed lighting on the Berlin TV Tower.
Microsoft's Execution Containers (MXC) are generally available on Windows 11, giving administrators per-agent file and network access policies with Windows enforcing hard isolation; isolation levels range from separate processes up to virtual machines and cloud PCs. HydraFusion enables models to run locally while accessing the cloud for additional scale, with GitHub collaboration in progress. Local agent apps including Muse, Hermes, OpenClaw, and Perplexity are being built for Windows, and Windows taskbar Search is adding thousands of quick actions this fall.
NVIDIA released Nemotron 3.5 Super VL, a 120B-parameter open model claimed to be five times faster than the next-best model in its class while matching frontier accuracy at one-fifth the parameter count of NVIDIA's largest reasoning model. NVIDIA released the majority of its training data alongside an in-depth tech report, and the model runs on consumer hardware including RTX Spark devices. It was trained on approximately 26 trillion tokens with post-training on roughly 3.8 million samples. Dozens of organizations including Scale AI, CrowdStrike, Atlassian, Salesforce, and Perplexity are post-training or integrating Super VL. TrainLoop post-trained the model with dermatology data from Nolla, which NVIDIA said became the first company approved for AI to write initial patient prescriptions.
Nemotron 3.5 Bolt 74B-A9B, released October 15, is a 74B-parameter, 9B-active-parameter MoE model using 2-bit quantization to fit into 24GB, available through llama.cpp, vLLM, Ollama, Unsloth, LM Studio, Hermes, and OpenClaw. Bolt E4B 4GB, launching October 23, targets conversational and physical agents on GeForce RTX and Jetson. All three Nemotron models ship with open weights, open datasets, and training recipes.
Microsoft demonstrated DeepSeek V4 Flash, a 284B-parameter open-weight model, running locally on a Surface Laptop Ultra after quantization to 1.6 bits, reducing its memory requirement to approximately 60GB. NVIDIA previewed DGX Station for Windows, powered by the GB300 Grace Blackwell Ultra Desktop Superchip, delivering 748GB of coherent memory and up to 20 petaFLOPS of FP4 compute. RTX Spark runs the full CUDA platform, enabling developers to move workloads from RTX Spark laptops to DGX Station without rewriting code.
Adobe and NVIDIA are delivering up to 2x performance improvements in Photoshop, Premiere, and Substance 3D for RTX Spark, available in beta. The first native PyTorch-CUDA build for Windows on Arm, Python 3.15, and a native Conda installer with over 2,000 packages released at launch; 97% of the top 15,000 Python wheels are ready. Llama.cpp added decision model support, which NVIDIA describes as allowing agents to classify requests and select next actions up to 200x faster than traditional reasoning loops; Llama Windows, a native Windows 11 system-tray app, also released, and Microsoft is adding llama.cpp as a backend to Windows ML. DLSS 5 with 3D-Guided Neural Rendering launched October 16, with a new 2x faster AI model debuting in NBA 2K27 that boosts frame rates by up to 1.4x; Hogwarts Legacy follows November 5.