Co-authored with Devang Aggarwal, and Srinivasa Deevi.
Snapdragon X2 Elite laptops are paired with the best in-class Hexagon NPU with long battery life, and Kilo Desktop by Anaconda gives that NPU real work to do. Kilo Desktop is one app for building with AI, where you can plan, write, test, and debug code. Connect it to Microsoft Foundry Local, which serves open models on the Snapdragon NPU, and your prompts and code can stay on your machine.
Anaconda’s native Windows on Snapdragon support brings Python, conda, and vetted packages to Snapdragon PCs. Kilo Desktop sets up conda environments for you inside the app, so the laptop is ready for Python work from the first prompt. Foundry Local handles the model, Kilo Desktop handles the project, and the two talk through a standard OpenAI-compatible API.
An NPU built for local models
The Hexagon NPU in Snapdragon X2 Elite has 80+ TOPS, up from 45 TOPS in the Snapdragon X Elite. The NPU runs model inference on its own silicon, which leaves the Oryon CPU cores free for your Python code, tests, and builds. Explore Snapdragon X2 Elite.
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.
Local AI on the NPU with Foundry Local
Foundry Local is an end-to-end local AI solution for building applications that run entirely on the user’s device. It provides native SDKs (C, C++, C#, JavaScript, Python, and Rust), a curated catalog of optimized models, and automatic hardware acceleration - all in a lightweight package (~20 MB). The compact size makes it easy to integrate into your application and distribute to end users. It also has an accompanying CLI, which allows you to persist an OpenAI compatible endpoint to connect into other applications such as Kilo Desktop.
Foundry Local’s model catalog lists 11 models built for the Qualcomm AI Engine (QNN), including Qwen2.5 Coder and DeepSeek R1 Distill builds.
Install Foundry Local and the latest Qualcomm NPU driver, then pick a model. When you run a model by its alias, Foundry Local picks the best hardware it finds, preferring the NPU over the GPU and CPU. To pin the NPU build, use its full model ID:
foundry run qwen2.5-coder-7b-instruct-qnn-npu:1
foundry run deepseek-r1-distill-qwen-7b-qnn-npu:2
The service port can change between sessions, so run foundry service status to get the current address. Requests to /v1/models on that address return the exact model IDs you’ll enter in Kilo Desktop.
- Qwen2.5 Coder 7B Instruct (4.7 GB): writing, explaining, and fixing code
- Qwen2.5 Coder 1.5B Instruct (1.3 GB): fast completions on lighter tasks
- DeepSeek R1 Distill Qwen 7B (5.3 GB): step-by-step reasoning about data and bugs
- DeepSeek R1 Distill Qwen 14B (9.8 GB): harder reasoning, if memory allows
Connect Foundry Local to Kilo Desktop
Kilo Desktop treats Foundry Local as a custom AI provider, so the NPU model shows up in the same model picker as every other model. Open the Add custom AI provider dialog and fill in four fields:
- Provider name: something you’ll recognize, such as Foundry Local NPU.
- API format: OpenAI Compatible.
- Base URL: the address from
foundry service status, ending in/v1(for example,http://127.0.0.1:<port>/v1). - API key: leave it blank. The field is optional and only needed when an endpoint requires authentication.
Under Models, add each model ID exactly as /v1/models returns it, and click Add. Check Reasoning for the DeepSeek R1 model, since it works through problems step by step before answering. Leave Images unchecked for these text-only models. Click Add AI Provider, and both models appear in the chat’s model picker.
With a Foundry Local model selected, your prompts go to the NPU on your laptop rather than to a cloud service.
Coding workflow: fix a failing test with Qwen2.5 Coder
Here’s a hypothetical example. A developer opens a Python project in Kilo Desktop, where a date-parsing function fails two of its pytest cases. They select Qwen2.5 Coder 7B from the model picker and ask it to read the failing tests, explain what the function gets wrong, and propose a fix.
The model points to a time-zone assumption in the parser and suggests a small patch. The developer reviews the diff, applies it, and asks the agent to run the tests in the project’s conda environment. When one case still fails, they paste the output back into the chat, and the model revises its fix.
Each turn stays on the laptop: Qwen2.5 Coder runs on the NPU, while pytest and the conda environment run on the CPU. A smaller model like this works best on focused tasks, such as one function, one test file, or one error message, so keep each request narrow.
Notebook workflow: explore a dataset with DeepSeek R1
Here’s a second hypothetical example. An analyst has a folder of CSV files on support tickets and wants to know why response times rose last quarter. In Kilo Desktop, they switch the model picker to DeepSeek R1 Distill Qwen 7B and ask it to plan the analysis.
Because it’s a reasoning model, DeepSeek lays out its steps before writing code: load the files, check for missing values, then compare response times by team and ticket type. The analyst asks it to turn that plan into a Python notebook using pandas from the project’s conda environment, and runs each cell as it’s written.
Python does the calculations, while the local model decides what to check next and explains what the numbers show. With the notebook open next to the chat, the analyst can question a result, ask for a different chart, or switch to Qwen2.5 Coder when a cell needs debugging. The ticket data never leaves the laptop.
Get started
To get started, download Kilo Desktop by Anaconda and Foundry Local, pick an NPU model, and give it a task from one of your own projects. The model runs on your Snapdragon NPU, and your code, notebooks, and environments live in Kilo Desktop, so the whole workflow stays on your laptop.
Related: NVIDIA RTX Spark and Kilo by Anaconda: New Device to Put Local AI to Work