Two numbers came off the same stage at Qualcomm's Snapdragon Summit, and between them they describe the 2027 earbud more honestly than the keynote did. One is a radio. The other is a context window.
The radio is micro-power Wi-Fi 6E, built into Snapdragon Sound Elite Gen 2 (opens in a new tab), the first audio platform Qualcomm has pitched at AI agents rather than headphones. It lets an earbud reach the cloud on its own, without tunneling through a paired phone. The context window belongs to Bonsai, a 2-billion-parameter vision-language model PrismML ran entirely on AR1 Gen 1 glasses from the same stage: a 1.7B language model squeezed to a single bit per weight, a 0.3B vision encoder at four bits, and room for about a thousand tokens at a time.
TL;DR
- Snapdragon Sound Elite Gen 2 puts micro-power Wi-Fi 6E in the earbud, so it can reach the cloud without a paired phone in the loop.
- PrismML's Bonsai fits a 2B vision-language model into glasses by quantizing the language half to one bit, roughly 0.43 GB of weights against 1.66 GB for the four-bit version.
- Bonsai's thousand-token window holds about four minutes of talk, so the on-device model can answer and describe but cannot remember.
- The 2027 split is set: the device answers, the uplink remembers, and the only fight left is over whose uplink the earbud uses.
Why does skipping the phone matter for an earbud?
Because the phone operating system has always been the earbud's tax collector. Background-audio limits, Bluetooth bandwidth ceilings, a companion app that has to be open and foregrounded to do anything interesting - every pendant and earbud pays that tax, and the phone maker sets the rate. A radio that talks straight to the cloud removes the intermediary. Plaud made the same move in a cruder form by dropping a 4G eSIM into its earbud case, so the buds record and answer with no phone nearby; Qualcomm has now done it in silicon, as Wi-Fi 6E, and wrapped it in a layer it calls Snapdragon Sound Apps and Agents. The message to Apple and Google is not subtle: the assistant on your ear does not have to be the one on your phone.
Qualcomm's own figures put the chip at twice the AI throughput of the S7 Gen 1 part it replaces, up to 40 percent less power, and a 30 percent smaller footprint, with support for camera-equipped buds so the thing on your ear can soon see as well as hear. This is the same wholesale play as the white-label glasses Qualcomm seeded earlier this year: ship the reference design, let a hundred brands ship the hardware, and stay the company every one of them depends on. The radio is the part that matters, because it decides who the earbud phones home to.
A thousand tokens is about four minutes. What can live in it?
A command interpreter and a scene describer, and almost nothing past that. The one-bit trick is what makes a 2B model fit on glasses at all: PrismML reports about 0.43 GB of language weights against 1.66 GB for the same model at four bits, and roughly 15 tokens a second against 7, measured on a 4 GB test board against a Qwen baseline it picked. Fitting the model was never the question the product cares about. The window is. A thousand tokens of context is a wake-word engine with opinions. It can take a command or name what the camera is pointed at, and hold one exchange while it does. It cannot hold a week, a relationship, or the plan you mentioned yesterday.
That lands exactly where the arithmetic on Meta's puck left things: decode rate is memory bandwidth divided by the weight bytes each token costs, and a device that can only spare a few hundred megabytes for weights and a few minutes for context is a front end, not a mind. Bonsai is that same arithmetic with a vendor's number finally attached, and on a lighter chip - AR1, the glasses-and-earbud class, not the puck-class Reality Elite.
The honest caveat is that every figure here is Qualcomm's or PrismML's, several of them design-simulation by Qualcomm's own footnote, none from a shipping product or an independent lab. One-bit quantization at 1.7B is a lossy regime, and nobody has shown what Bonsai's accuracy does on a real scene task, let alone what the radio and the NPU together do to a battery the size of an earbud's. The numbers mark a ceiling a vendor drew for a demo, and no independent lab has had to redraw it.
So whose uplink is it?
The device answers and the uplink remembers - that part is settled for 2027. The memory, the planning, the assistant that writes a tool call before it says a word, all of it lives across the link, because the arithmetic will not let it live on the earbud. Whoever owns the link owns the user. There are two candidates. The phone maker, whose operating system used to be unavoidable and still owns the screen where a purchase gets confirmed. And the chip, which just made the phone avoidable and would very much like the earbud to answer to a cloud agent of the buyer's choosing instead.
Who loses if the direct radio wins is the easier call: every pendant and companion-app earbud whose whole architecture assumed the phone would sit in the middle, holding the microphone session open and brokering the cloud round-trip. Take the phone out and that design has nothing left to sell but the strap.
Two dated calls. The first Snapdragon Sound Elite Gen 2 earbuds to ship will carry a cloud assistant that is not the phone maker's, by the end of Q2 2027. And no AR1-class glasses will ship an on-device model with a context window over 4,096 tokens during 2027 - the window, not the parameter count, is the ceiling the hardware enforces, and a thousand tokens is where this generation sits.
The reference design for an always-reaching earbud now exists, and the model meant to run beside it is small enough to prove most of the work happens somewhere else. What is still up for grabs is the somewhere else.