Unsloth Desktop has launched. It bills itself as the first desktop application that lets you run and train models locally.

Its core features are straightforward: it allows you to connect Claude Code and Codex to a local LLM; it claims 2x faster training speed and a 70% reduction in VRAM usage; it includes built-in private web search and RAG. It is open-source. Its tagline reads: "The cloud has become an option."

The video showcases the app's interface: the left sidebar holds modules for model center, projects, training and more, while the right side is a chat window that supports file uploads, web search, code, deep research, and even MCP. The interface looks just like a regular chat app, but under the hood, it runs your own model.

Some users who have tested the app reported that their GPU fans spin extremely loudly, like the desktop is about to take off, but local training is indeed fast. Other users noted that the line "the cloud is now an option" is a bold claim, since training still requires a powerful GPU.

The most immediate benefit of moving training from the cloud back to local machines is that data never leaves your device. This is an essential requirement for developers working with sensitive data. RAG and private search also mean you can connect your internal documents to the model without routing data through third-party APIs.

Of course, a 70% reduction in VRAM usage does not mean there is no VRAM requirement at all. The barrier to entry for local training still exists, but it is getting lower. Compatibility also needs to be tested in real-world use.

"The cloud has become an option" — this does not mean the cloud will disappear, but rather that it is no longer the only choice. For developers, having an extra option is always a good thing.