CodeGuide teases Mac-1, a local 6.6B model built for macOS tools
Zafir says Mac-1 runs on Macs with 8GB-plus RAM and can chain tasks across 487 native macOS tools at about 65 tokens per second.
By Ryan Merket ยท Published
Why it matters
Mac-1 reflects a broader shift from cloud chatbots toward local desktop agents, where latency, privacy, and OS-level tool access matter as much as model size.
CJ Zafir (@cjzafir) teased Mac-1, a 6.6B-parameter model that he says runs locally on Macs and is built to operate native macOS tools, in a seven-post thread on X on Saturday.
https://x.com/cjzafir/status/2063305526285578573
Zafir claims Mac-1 requires 7GB of RAM, runs best with 12GB, and produces roughly 65 tokens per second with reasoning turned on. In a reply, he clarified that the target is Macs with 8GB or more of RAM, narrowing the "any Mac" claim in the original post.
The important part is not just the model size. Zafir says his team built an application layer around Mac-1 so it can use 487 native macOS tools and perform multi-tool chained tasks. That places the project closer to a local desktop agent than a general chatbot, with the operating system itself as the action surface.
Zafir framed the target explicitly in replies: "Better UX than Raycast," he wrote to Zach (@zach_sndr), and answered David Walker (@walkering420)'s prompt with "Mac1 vs Siri." Those comparisons put Mac-1 against two different incumbents: Raycast as the power-user command layer on macOS, and Siri as Apple's default assistant.
The thread did not specify a release date, pricing, license, benchmark methodology, or whether Mac-1 will ship as a downloadable model, a Mac app, or both. For now, the claims are Zafir's own demo posture: small enough to run locally, wired deeply enough into macOS to do useful work, and positioned against the assistants Mac users already know.