A practical guide to local models

Open Source Offline AI That Works Without the Internet

Cursiv is MIT-licensed desktop software that can run compatible language models locally through Ollama. Local inference means the model generates its answer on your computer; it does not, by itself, guarantee that every app feature works offline or that every model is open source.

Cursiv is in active development. Answers can be wrong; verify important information.

Local AI, offline AI and open source mean different things

Local inference means a model runs on your computer using a local runtime. Your prompt is sent to that runtime. The model and its files need to be present on the device.

Offline use means a feature can operate without an internet connection. Cursiv's local chat and local data features can work offline after setup, provided the compatible model and required files are already available. Setup downloads, web search, online updates and cloud providers require connectivity.

Open-source software describes Cursiv's code and license: its project code is released under the MIT License. That does not make every model's weights or license open source. Model terms vary, so check the license for the particular model you download. See model names and license notes.

What can work offline?

Once Ollama and a compatible model are installed, local chat can generate responses without contacting an online AI provider. Cursiv can also use locally stored conversations and memory, its included offline reference, imported documents and built-in calculations, depending on the feature and files available.

Cursiv includes an offline mode that blocks connections from the application to outside services. It remains enabled across restarts until turned off. The published offline demonstration reports 15 of 15 checks passing with no outside connection attempts; this is evidence about that test procedure and build, not a guarantee about every computer or a security certification.

Read the offline demonstration · Read the detailed privacy notes

Features that need a connection

Installing or downloading a model, browsing or searching the web, online update checks, optional cloud AI providers and Cursiv Cloud need internet access. A provider receives the prompt or content sent to it when you choose that integration. The phone app is a website feature and requires connectivity.

When a local model answers, that inference is local. Local storage is not the same as disk encryption: Cursiv documents that conversations and memories are not encrypted on disk. Saved AI keys have separate platform-specific protection.

Offline AI for field research

Cursiv's desktop Field Notebook records observations, site and sample details, coordinates, notes and photos on the computer. Entries preserve measurements as typed; corrections are added as new entries with a reason. The record chain is tamper-evident, which can reveal later edits, but it does not prove that an observation is true. The project documents export to CSV, JSON and GeoJSON.

The companion phone notebook can capture records without a signal after it has first been opened and saved on the phone. It does not run Cursiv's AI model on the phone. It stores entries in the browser and exports a file for transfer to the desktop app. The phone workflow has not been tested on real iPhones or Android phones, and browser storage can be cleared, so the project recommends sending records to the laptop at the end of each day. Try the phone field notebook and read its status and limits.

Models, runtime and hardware

Cursiv uses Ollama to run compatible models locally. The project currently recommends Qwen3 4B Instruct as its default on supported setups; a smaller Qwen2.5 1.5B option is also listed. These models have their own licenses, which are separate from Cursiv's MIT license.

The README lists Windows 10 or 11 (64-bit), 8 GB of system RAM and about 6 GB of free disk space for one model as starting requirements. A graphics card is not mandatory: CPU inference is supported, usually at a different speed. Available memory, model size, quantization, context length and other running programs affect whether a model fits and how responsive it feels.

For model-by-model license details and hardware measurements, consult MODELS.md and the measured hardware notes. These measurements are from one computer and should not be treated as a universal speed promise. See the local AI hardware guide for choosing a suitable setup.

Install Cursiv

  1. Download the Windows installerGet the current release from the download page.
  2. Install and create a local loginThe setup window guides the initial configuration.
  3. Install Ollama and a modelThis download needs internet access. Once the model is installed, try local chat and enable offline mode when you need to block outside connections.

Frequently asked questions

Does Cursiv work without internet?

Supported local features, including chat with a compatible installed model, can work offline after setup. Downloads, web search, cloud integrations, online updates and the website-based phone app need a connection.

Does local AI mean my data is encrypted?

No. Local inference keeps that model interaction on the computer, but local storage does not automatically mean encryption. Cursiv says conversations and memories are not encrypted on disk. Review the privacy documentation for details.

Are the AI models open source?

Cursiv's own code is MIT-licensed. Models downloaded through Ollama have separate terms; some are permissively licensed and others have additional conditions. Check the exact model's license before use.

Can Cursiv use cloud AI too?

Optional cloud providers are available in supported configurations. They require internet access and send the content needed for that request to the selected provider. Offline mode blocks outside connections.

What computer do I need?

The project README lists Windows 10/11 64-bit, 8 GB RAM and roughly 6 GB free disk space for one model as a starting point. Model and speed fit vary. See the hardware guide.