On 16 October NVIDIA published a developer note on RTX Spark. The machines use the GB10 chip, up to 128GB of unified memory, and 1 petaflop of FP4 AI compute. Laptops are available now. Small desktops are described as arriving in November, for agents that stay on.

[1]
A compact desktop computer with a charcoal case and a muted-blue side
A compact desktop computer, the front of the case painted charcoal and one side muted blue, with a small round port., AI-generated illustration, not a news photograph

The speed figures are NVIDIA’s own comparison, not a third-party rerun. Against Intel Panther Lake 12Xe, coding-agent performance is 4.6 times, fine-tuning 2.7 times, and code compilation 1.8 times. The figure caption sets an Intel Core Ultra X9 388H to 1.0. The three tests are AA-AgentPerf-local with Qwen3.8-27B, Unsloth LoRA with Llama-3.1-8B, and a Linux kernel compile in the Phoronix suite. The name in the prose and the chip in the caption both stay. They are not collapsed into one part number.

[1]

The same post announces Nemotron 3.5 Bolt 74B-A9B 24GB. It is a 74-billion-parameter mixture-of-experts model with 9 billion active parameters, packed into 24GB with 2-bit quantization. NVIDIA says it is meant to balance speed and capability, and that at launch it will run through llama.cpp, vLLM, Ollama, Unsloth, and LM Studio. Weights, training data, and recipes are described as open, as with other Nemotron models.

[1]

Decision models are described as choosing a next step up to 200 times faster than traditional agents, for classifying requests and selecting actions, and as available in llama.cpp and Unsloth. The 200-times figure is NVIDIA’s upper claim. This story has no separate measurement. The same partnership note also gives a first PyTorch-CUDA build on Windows on Arm, Python 3.15 native by default, more than 2,000 Conda packages, and Arm builds for more than 97% of the top 15,000 Python packages.

[1]