NVIDIA RTX Spark brings powerful local AI computing to Windows laptops, including the new Microsoft Surface Laptop.
RTX Spark combines NVIDIA RTX graphics, the full NVIDIA AI stack and up to 128GB of unified memory for developing, creating, and gaming in one device. Kilo Desktop by Anaconda puts that power to work. Kilo Desktop is one app for building with AI, where you can plan, write, test, and debug code. Its built-in local model server runs open models on your RTX hardware, so your prompts can stay on your machine.
Anaconda’s native Windows on Arm support brings Python, conda, and vetted packages to the platform. RTX Spark runs the full NVIDIA CUDA platform, and this is the same software stack that runs across NVIDIA hardware, which Anaconda supports with Python artifacts.
Kilo Desktop sets up conda environments for you inside the app, so your RTX Spark device is ready for Python work from the first prompt. Start with a real task: explore a dataset, fix a failing script, or build an assistant that searches your project files. The local model and your tools all work in the same app.
Hardware with room for local models
RTX Spark pairs NVIDIA Grace CPU cores with a Blackwell RTX GPU, and some configurations include up to 128 GB of unified memory. Because the CPU and GPU share that memory, you have room for larger models and more of the context they need while they work. NVIDIA also brings native CUDA support to the platform. Explore RTX Spark.
On the software side, Anaconda’s Windows on Arm support provides native Python, 2500+ conda packages, and vetted libraries. In Kilo Desktop, you manage a project’s environments in the same app where you work with the model, so your code, notebooks, and data stay together.
Bring Nemotron into the project
The NVIDIA Nemotron family includes open models built for reasoning and tool use, which matter when an agent has to work through several steps. A suitable Nemotron model can interpret a request, decide which tool to call, and use the result to choose its next action. Explore Nemotron examples.
Pick a model version that fits your RTX-based machine’s memory, and import GGUF model files directly into Kilo Desktop’s built-in server to add them to the agent’s model picker. For agent work, choose a model that supports tool calling, then turn on coding tools and instructions in its settings. Local model settings.
With local inference selected, your prompts are processed on your device, keeping your workflow local.
Build an agent around a real question
Here’s a hypothetical example. A product team wants to know why returns went up, and it starts with a folder of CSV files. In Kilo Desktop, a developer could ask the agent to inspect the columns, flag missing values, and write a Python notebook that compares return rates by product and month.
The developer reviews the code and results, then asks the agent to look into the largest changes. Python does the calculations, while the local model directs the work and explains what the numbers show. With the notebook open next to the chat, the analysis is easy to check and revise.
The same approach works for a document assistant, a coding agent that runs your project’s tests, or a workflow that turns new local files into a report. Each one starts with a specific task and the tools the model needs to do it, and you can run the whole development process locally on Windows PCs powered by NVIDIA RTX Spark.
Get a first local response in under five minutes
Once Kilo Desktop, the model file, and your project’s packages are installed, the goal is a first local response in under five minutes. The initial downloads and setup take longer, depending on your connection and the model you choose.
- Open Kilo Desktop, sign in, and select your project folder.
- Select the project’s conda environment, or import its
environment.ymlfile. - Import a compatible Nemotron model, start the model server, and select the model in chat.
- Turn on coding tools for a model that supports them, then give it a small task, such as explaining a Python file.
That first response gives you a baseline. From there, add tools and try larger tasks as you check the model’s results. Kilo’s local inference guide covers importing and selecting a model.
Keep the setup with the work
As the project grows, save its Python requirements in an environment.yml file. Record the model version and settings next to the code so teammates with compatible hardware can recreate the workflow. Kilo Desktop can create and import conda environments, which keeps those dependencies with the project.
To get started, download Kilo Desktop by Anaconda, pick a compatible local model, and give it a task from one of your own projects. The model runs on your RTX Spark, and your code, notebooks, and environments live in Kilo Desktop.