Get the newsletter
Qualcomm · AI / ML · Filed May 16, 2025 · USPTO published Nov 19, 2026 · This week's release

A New Qualcomm Patent Lets Several Devices Share the Work of One AI Model

Qualcomm has filed to split one AI model across several devices, so your phone, laptop and tablet could each carry a slice of the load.

An AI program divided into sections so multiple phones or computers can run parts of it together. Drawing from patent filing US 2026/0368485 A1.
An AI program divided into sections so multiple phones or computers can run parts of it together.All 15 drawings ↓
Publication number
US 2026/0368485 A1(US20260368485A1)
Applicant
QUALCOMM Incorporated
Filed · Published
May 16, 2025 · Nov 19, 2026
Inventors
Qi XUE, Abhijit NAVALEKAR
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/15

What Qualcomm's device-sharing AI patent does

Imagine you ask an AI assistant a big question, but your phone is too small to run the whole model by itself. Normally the answer is to send the job to a data center. Qualcomm's patent application describes another route: let a few devices split the job between them.

Your phone, a laptop and a tablet would each take a slice of every step in the model. After each step, they swap results so everyone has the full picture before the next one starts. It works a bit like a few people assembling one jigsaw puzzle, each handling a section and then trading pieces.

Slice sizes vary. A stronger device gets a bigger slice, so all the devices finish at about the same time. The filing also describes a way to pick between two splitting styles depending on which is faster for a given job.

From the filing · CLAIM 1
process, using a first portion of a first layer of a machine learning network, an input embedding to generate a first output corresponding to the first portion of the first layer of the machine learning network …

Translation: Your device handles just one specific slice of the AI network's first layer.

How each device handles its slice of a layer

AI models are built as a stack of layers, and each layer is a huge grid of stored numbers (called weights). Running a big model means pushing your input, turned into a list of numbers called an input embedding, through every layer in turn. The patent covers cutting each layer into portions, one per device. Engineers call this tensor parallelism (splitting the number grids themselves across machines).

Here is the loop the claims describe:

  • Each device runs the input through its own portion of the first layer.
  • One device gathers the other devices' results and combines them into the full output of that layer.
  • That combined output is sent to the other devices.
  • Every device then runs its portion of the second layer on it, and the pattern repeats for the third.

Dependent claims add details. The combined result can go out one device at a time (called unicast, meaning one-to-one messages), and the order can flip on the next round. Slice sizes can follow each device's processing power, and the layer can be a self-attention layer (the part that weighs which words in a prompt matter to each other).

The description adds a second idea: use pipeline splitting (each device owns a block of whole layers) while the model reads your prompt, then tensor splitting while it writes the reply. A lookup table of speed measurements helps choose. In its example, three devices handle the first stage and two handle the second.

From the filing · THE ABSTRACT
… determine a first combined output embedding based on a combination of the first output and a plurality of outputs corresponding to additional portions of the first layer of the machine learning network received from one or more additional devices …

Translation: The system merges your device's partial results with data sent from other nearby gadgets.

Why this matters for AI on phones and laptops

The filing starts from a familiar problem: phones, tablets and laptops have limited power, memory and battery, so AI models either get shrunk or get sent to cloud servers. Splitting one model across devices you already own offers a third option, one that could let bigger models run on gear in your own hands.

There is a catch the filing itself points to. The devices have to trade results after every layer, so a slow connection between them eats into the gains. That is why the description spends so much time on link speed, device count and a lookup table for choosing the fastest setup. The figures show a chip with a neural processing unit, 5G, Wi-Fi and Bluetooth, which fits a phone-class setting.

This is the 70th Qualcomm filing we've tracked since July in the AI chip wars, building on earlier applications like one splitting a model across devices and one for a phrase cheat sheet.

Source: US 2026/0368485 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Qualcomm patent news breaks.

There are more where this came from

We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.

The drawings · 15 sheets

15 drawing sheets from US 2026/0368485 A1 · click any drawing to enlarge

Main processor parts on the chip that run everyday tasks, drawing sheet 1 from US 2026/0368485 A1
Sheet 1 Main processor parts on the chip that run everyday tasks
Slicing an AI program so several gadgets can share it, drawing sheet 2 from US 2026/0368485 A1
Sheet 2 Slicing an AI program so several gadgets can share it
Passing whole tasks sequentially down a line of devices, drawing sheet 3 from US 2026/0368485 A1
Sheet 3 Passing whole tasks sequentially down a line of devices
Splitting tasks into smaller pieces to speed up responses, drawing sheet 4 from US 2026/0368485 A1
Sheet 4 Splitting tasks into smaller pieces to speed up responses
Exchanging data between devices running the same AI model, drawing sheet 5 from US 2026/0368485 A1
Sheet 5 Exchanging data between devices running the same AI model
Benchmark speeds used to pick the fastest processing setup, drawing sheet 6 from US 2026/0368485 A1
Sheet 6 Benchmark speeds used to pick the fastest processing setup
Picking the ideal device count to minimize total wait time, drawing sheet 7 from US 2026/0368485 A1
Sheet 7 Picking the ideal device count to minimize total wait time
Tuning device counts separately for reading prompts and generating text, drawing sheet 8 from US 2026/0368485 A1
Sheet 8 Tuning device counts separately for reading prompts and generating text
Show the other 7 sheets
Three gadgets take turns before two split the reply, drawing sheet 9 from US 2026/0368485 A1
Sheet 9 Three gadgets take turns before two split the reply
How several devices pool results to finish an AI task, drawing sheet 10 from US 2026/0368485 A1
Sheet 10 How several devices pool results to finish an AI task
Switching between relay and team modes to generate answers, drawing sheet 11 from US 2026/0368485 A1
Sheet 11 Switching between relay and team modes to generate answers
A standard deep learning setup that finds patterns in data, drawing sheet 12 from US 2026/0368485 A1
Sheet 12 A standard deep learning setup that finds patterns in data
An image scanner that helps an AI recognize street views, drawing sheet 13 from US 2026/0368485 A1
Sheet 13 An image scanner that helps an AI recognize street views
The text generator that turns prompts into finished replies, drawing sheet 14 from US 2026/0368485 A1
Sheet 14 The text generator that turns prompts into finished replies
Internal computer parts that run the shared AI tasks, drawing sheet 15 from US 2026/0368485 A1
Sheet 15 Internal computer parts that run the shared AI tasks
Patent filing page

More in AI training & infrastructure

All 323 AI training & infrastructure patents →

Watchlists this filing belongs to

All watchlists →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.