Big Models, Small Footprint: Local AI with Kilo Desktop Anaconda's Kilo Desktop now provides a full local AI development stack on the Dell Pro Precision with GB10, a desktop workstation built on the NVIDIA GB10 Grace Blackwell Superchip with 128 GB of unified memory. Kilo Desktop lets developers manage conda environments, download and serve open-weights models through an OpenAI-compatible local model server, and write code and notebooks with an AI agent running on those local models, with Anaconda building Python, conda, and its repository packages natively for Arm64 Linux. The setup targets teams that previously rented cloud GPUs or settled for smaller models, keeping prompts, responses, and underlying data on the device once a model is downloaded. Some AI models are simply too big for a personal machine. Without specialized hardware, teams have needed to rent cloud GPUs and wait in a shared cluster queue, or settle for a smaller model. Dell Pro Precision with GB10 https://www.dell.com/en-us/shop/workstation/spd/dellpromaxwithgb10fcm1253/fcm1253 reg 01 changes that math. Built on the NVIDIA GB10 Grace Blackwell Superchip with 128 GB of unified memory, it fits on a desk and runs models that used to need a server rack. Powerful hardware is only the start of a working setup. Developers need Python, packages that are built for the chip, a place to serve models, and a way to write code against them. Anaconda and Kilo Desktop now cover the development stack on the Dell Pro Precision with GB10 https://www.dell.com/en-us/shop/workstation/spd/dellpromaxwithgb10fcm1253/fcm1253 reg 01 , so the machine is ready for AI work the day it arrives. Kilo Desktop https://kilo.ai/install is Anaconda's agentic engineering app for developers and data scientists who build AI applications and analyze data. It puts the pieces of local AI work in one place: you manage conda environments and Anaconda packages https://www.anaconda.com/products/packages-and-models , download and serve open-weights models https://kilo.ai/docs/desktop/features/local-inference , and write code and notebooks with an AI agent https://kilo.ai/docs/desktop/features/notebooks that runs on those local models. There has never been a better time to be a builder. A full stack built for the chip The GB10 pairs an Arm-based Grace CPU with a Blackwell GPU, and both share one pool of memory. That design is what lets a single desktop hold a large model, but it only pays off when the software above it is built for the same architecture. A package compiled for a different processor either won't install or leaves performance on the table. Anaconda builds Python, conda, and the packages in its trusted repository https://www.anaconda.com/products/packages-and-models natively for Arm64 Linux, the platform the GB10 runs on. When a data scientist installs NumPy, pandas, PyTorch, or scikit-learn from Anaconda, they get a build made for this hardware and vetted by Anaconda, with nothing pulled from an unknown source. Environments you can see and share Every AI project carries its own set of dependencies. Unfortunately, they rarely agree with each other. A fine-tuning experiment may need one PyTorch version while a forecasting notebook requires another. Conda keeps them apart in separate environments https://kilo.ai/docs/desktop/features/environments , and Kilo Desktop gives those environments a home you can see. In Kilo Desktop, you can create an environment https://kilo.ai/docs/desktop/features/environments , add Anaconda packages, or import an existing environment.yml file without opening a terminal. The same file rebuilds the same environment on a teammate's laptop, a second GB10 system, or a cloud server. A project can start on the desk and move elsewhere later without anyone reverse-engineering what was installed. Larger models, served from your desk The GB10 can hold open-weights models far larger than a typical workstation GPU allows. That makes available open models that are better at reasoning, coding, and parsing long documents without sending a single prompt off the machine. Kilo Desktop handles the serving. Pick an open-weights model, download it, and start a local model server https://kilo.ai/docs/desktop/settings/ai local-model-server from the app. The server speaks the OpenAI-compatible API, so code written against a hosted model can switch to the local one by changing its base URL. Once the model is downloaded, prompts, responses, and the data behind them stay on the device. An AI partner in the notebook, running locally Serving a model is useful on its own, but the bigger change comes when that model works alongside you. Kilo Desktop includes agentic coding https://blog.kilo.ai/p/kilo-desktop-is-generally-available and notebook tools https://kilo.ai/docs/desktop/features/notebooks that run on the models you serve locally. You describe what you want load this dataset, clean these columns, chart revenue by region , and the agent writes and runs the cells inside your conda environment, so it uses the packages you chose. Because the agent and the model both run on the GB10, you can point it at sensitive data https://www.anaconda.com/blog/nvidia-rtx-spark-kilo-desktop-local-ai that could never go to a hosted assistant. There are no per-token charges to watch, and you can let the agent iterate as many times as the task needs. Your first session on the Dell Pro Precision with GB10 Here is what a first session may look like, from power-on to a working analysis: 1. Install Kilo Desktop https://kilo.ai/install and connect it to Anaconda's vetted package repository https://www.anaconda.com/products/packages-and-models . 2. Import your project's https://kilo.ai/docs/desktop/features/environments environment.yml , or build a new environment in the app. 3. Download an open-weights model and start serving it locally https://kilo.ai/docs/desktop/features/local-inference from Kilo Desktop. 4. Open a notebook https://kilo.ai/docs/desktop/features/notebooks , attach the environment and the local model, and ask the agent to explore your data. 5. Turn the result into something others can use, such as a Streamlit app that answers questions about the dataset with the same local model. Find your perfect use case Data scientists and AI developers get a machine that can run serious models without a cloud account or a cluster ticket. It suits experiments that need a large model, work on regulated or private data, and fast iteration where waiting on remote GPUs would slow you down. Teams and IT https://kilo.ai/teams get reproducibility and control. Environment files keep setups identical across machines, Anaconda's repository gives them a vetted source for every package, and local inference https://kilo.ai/docs/desktop/features/local-inference keeps sensitive data inside their walls. The Arm64 Linux ecosystem gains the Python tools most data scientists already use, built natively and ready on hardware designed for AI. Get started Anaconda and Kilo Desktop are available now for Dell Pro Precision with GB10.