New AI Tech 'Nexus' Makes Device Assistants 1.6x Faster

AI assistants are getting more useful — they can now check your calendar, search files, or send messages by using built-in “tools.” But every time you ask something, the assistant has to re-read the instruction manual for every tool it knows. With more tools, this slows things down and eats up memory. That’s the problem a new research paper tackles.

The solution, called Nexus, takes a different approach. Instead of plowing through every tool description, it uses a lightweight “cheat sheet” to quickly pick the right tool. Then it generates the answer using a short compressed version of that tool — about 19 words on average, rather than the full manual. This means far less reading, so the assistant can get to work faster. The system acts like a quick-reference guide while the full manuals stay in storage until truly needed.

In testing on an Apple Silicon Mac with a 14-billion-parameter AI model, Nexus produced the first meaningful response up to 1.66 times faster than the old method. It also saved about 80% of the main memory space. Even with a huge toolbox of 250 tools — a scenario that would crash a normal system — Nexus kept its accuracy at nearly 89%. That’s a big deal for anyone who wants a powerful assistant on a personal device rather than in the cloud.

There are honest limits. The speed boost shrinks during very long conversations, sometimes dropping to no improvement. Accuracy isn’t perfect, and the numbers come from one specific lab setup, so real-world results may vary. But the paper shows a clear path toward faster, lighter AI agents on everyday hardware. For users, that could mean voice assistants that respond without awkward delays, and AI features that work smoothly on a laptop without draining battery or relying on an internet connection.