Products·2 min read·TechCrunch

Osaurus Lets Mac Users Mix Local and Cloud AI Models in One Sandboxed App

Built by ex-Tesla and Netflix engineers, Osaurus routes between local Llama, DeepSeek V4, and Gemma 4 and cloud GPT-5.5, Claude, and Gemini in one sandboxed macOS app — with 112K downloads so far.

Share:

A pair of ex-Tesla and Netflix engineers have shipped Osaurus, an open-source macOS app that finally treats local and cloud AI models as interchangeable rather than as separate worlds. The app lets a user flip between an on-device Gemma 4 running on their own M-series chip and a cloud-hosted Claude or GPT-5.5 within the same conversation, while keeping the user's files, memory, and tool calls scoped to their own hardware via a sandboxed virtual machine.

Co-founders Terence Pae and Sam Yoo originally built a desktop AI companion called Dinoki, then rebuilt it as Osaurus once it became obvious that the bigger problem for Mac users was not which model to run, but how to compose models, files, and apps without leaking data to a half-dozen vendors. The current build ships with more than 20 native plugins — Mail, Calendar, Vision, Browser, Git, and others — alongside a full Model Context Protocol server, so anything that speaks MCP can plug straight in. Local model support covers MiniMax M2.5, Qwen3.6, GPT-OSS, Llama, DeepSeek V4, Apple's on-device models, and Gemma 4; cloud routing handles OpenAI, Anthropic, Gemini, xAI/Grok, and OpenRouter behind one interface.

The pitch is positioning rather than benchmark. Osaurus is not trying to beat Ollama or LM Studio on raw model throughput; it is trying to be the layer above them — an ordinary-user shell where switching between local privacy and cloud capability is one click, and where the integrations matter as much as the model. The hardware bar is steep: 64GB of unified memory minimum, with 128GB recommended once users start loading DeepSeek V4. That positions Osaurus squarely at the new tier of M-class Macs that Apple has been pushing into developer hands ahead of WWDC in June.

Adoption is already non-trivial. The app has been downloaded more than 112,000 times since its release just under a year ago — modest by App Store standards, but extremely high for a developer-skewed local-AI tool. That gives Pae and Yoo a beachhead at exactly the moment Apple is widely expected to use WWDC 2026 to reveal a Google-Gemini-powered Siri overhaul and a deeper opening of its on-device AI stack. If Apple commodifies cloud-routing into the OS, Osaurus has to climb up the stack. If Apple doesn't, the gap it is filling becomes bigger, not smaller.

Comments

Share your thoughts. Be kind.

0/2000

Loading comments…

Related Articles