New AI on Mac Remembers a Million Words at Once

Most AI chatbots today remember only a few pages of conversation before they start forgetting. That's because they rely on a limited working memory called "context." This new release changes the game for people with very powerful Macs: it lets Qwen3.8-AI keep a million tokens in mind at once. That's roughly like feeding it “Winnie-the-Pooh” and expecting it to quote any line later. The person behind it built this not as a demo trick, but as a serious tool for real long-form work.

The magic is in the hardware and software. Instead of needing massive cloud servers, it runs on a Mac with Apple's M5 Max chip and 128GB of memory, using a tool called MLX-serve. To fit such a giant memory, the AI's brain is compressed smartly: core parts are kept in 8-bit precision and expert parts in 4-bit, so it stays sharp without breaking your laptop. On early tests, it keeps producing about 40 words per second on normal text and even faster on coding tasks. That's fast enough to generate a long essay while you watch.

So why should you care? Because it moves powerful AI out of giant data centers and onto a personal device. That means you can analyze an entire book, a massive legal document, or whole software project without sending your data to a cloud company. For anyone working with confidential or enormous texts, privacy and convenience gain enormously.

The honest catch: this is early software for developers. You need roughly 117GB of free memory, a Mac that expensive hobbyists own, and comfort using command-line tools. There will be bugs, and normal users won't use it today. But it's a clear sign that super-long-memory AI is coming to your own computer—not just to billionaires' servers.