On September 23, PrismML announced a demonstration of "Bonsai," a roughly 2-billion-parameter vision-language model, running entirely on-device on smart glasses equipped with Qualcomm's Snapdragon AR1 Gen 1. The language processing portion is configured in 1-bit, while the image processing portion is configured in 4-bit.
The intended use case is for glasses to answer questions on the spot about what they're looking at. However, what was shown this time is a technical demonstration—no commercial product launch featuring Bonsai, nor a timeline for consumer availability, has been announced.
How was the model shrunk down to run on smart glasses with only limited memory and power available, and to what extent was its performance verified? Understanding these two points separately is key to interpreting this announcement.
What ran on the smart glasses was a model combining language and vision
A parameter count of roughly 2 billion does not mean the entire model is uniformly a "1-bit AI."
According to PrismML's announcement, the 1.7-billion-parameter language model is represented in 1-bit, while the roughly 300-million-parameter image encoder is represented in 4-bit. Combining these two creates a vision-language model capable of simultaneously handling images captured by the camera and language.
In other words, the image processing portion was not compressed down to 1-bit as well.
PrismML and Qualcomm Technologies state that they optimized the model's weights and architecture to match Qualcomm's Hexagon NPU.
The NPU is a computation circuit specialized for AI processing, and the Snapdragon AR1 Gen 1 also incorporates functionality for processing video from the camera.
For AI that handles visual information, it's not enough to simply shrink the language model that generates text. The processing required to pass images captured by the camera to the model, as well as the portion that extracts features from images, must also fit within the smart glasses' limited memory and power budget.
For this Bonsai demonstration, the language portion was set to 1-bit and the image portion to 4-bit to achieve this.
This is not the first demonstration of generative AI running standalone on smart glasses.
In June 2025, Qualcomm reported running an AI assistant using Llama 1B on the RayNeo X3 Pro, equipped with Snapdragon AR1+, without relying on a smartphone or the cloud.
What's different this time is that PrismML used a model optimized for the AR1 Gen 1 and ran a roughly 2-billion-parameter vision-language model combining language and images.
However, since the AR1+ and AR1 Gen 1 are different chips, one cannot simply compare speed differences between generations based on these two demonstrations.
What the "4x" and "2x" figures actually indicate
PrismML explains that, for the same amount of memory, it may be possible to fit a model with up to 4 times more parameters than before.
However, this "4x" figure does not represent performance that is guaranteed under general conditions. The company's own wording limits this possibility to "some smart glasses designs."
Looking at the measured values listed in the footnotes of the announcement materials, the comparison scope narrows further.
| Published comparison | 1-bit language model | Corresponding 4-bit language model | Difference |
|---|---|---|---|
| Memory occupied by 1.7B language model weights | 0.43GB | 1.66GB | 3.83x per Qualcomm measurement, ~74% reduction |
| Token generation speed | 15.36 tokens/sec | 7.44 tokens/sec | 2.06x per Qualcomm measurement |
The "4x" figure in the announcement is an explanation about the model scale that can fit within the same amount of memory. Meanwhile, the measured 3.83x difference compares the memory footprint of the language model's weights, and the 2.06x figure compares text generation speed.
Measurements were conducted under the conditions of Snapdragon AR1 Gen 1, 4GB of memory, and a context length of 1024 tokens, using a model compiled with an internal QNN SDK build that supports 1-bit computation.
Note that 0.43GB and 1.66GB are rounded published figures. Therefore, one should not simply divide these two numbers to extract precision beyond the 3.83x figure that PrismML presented.
Also, the 0.43GB figure represents only the capacity occupied by the language model's weights.
On actual smart glasses, memory is also needed for the image encoder, input data, working memory used during inference, the OS, and more. The 4GB figure represents the total memory capacity of the platform used in testing—it does not mean "the entire model ran within 0.43GB."
Being able to store the model's weights and being able to comfortably run the entire system, including vision AI, are two things that need to be considered separately.
The same applies to speed.
The figure of 15.36 tokens/sec represents the speed of the portion where the language model generates text sequentially. It has not been disclosed how many seconds it takes to capture a camera image, extract features from the image, and display the first response.
In other words, this figure alone does not reveal how long a user would wait between asking a question and receiving an answer.
While Qualcomm notes that 15.36 tokens/sec is a conservative figure, there's no guarantee the same speed would be achieved with longer inputs or on different smart glasses.
For a product worn for long periods on a daily basis, battery consumption and heat generation also need to be measured continuously, not just processing speed.
The context length of 1024 tokens set during measurement is also merely a test condition for this occasion. It is not a specification indicating the maximum context length of a production version.
How required memory capacity and processing time change during extended conversations or when referencing multiple images needs to be confirmed separately.
PrismML also explains that, on the language benchmarks cited in the announcement, the 1-bit version showed performance close to the 4-bit version.
However, the basis shown in the footnote for this is an evaluation comparing the company's own 1.7B 1-bit language model against the 4-bit version of Qwen 3 1.7B.
While the names of eight benchmarks that assess code generation, knowledge, and instruction-following—such as BFCL v3 and HumanEval+—are listed, specific scores for each are not provided.
Moreover, these results do not reveal the accuracy rate for correctly answering questions about images captured by smart glasses, nor do they confirm the overall performance of the roughly 2-billion-parameter vision-language model as a whole.
Conditions that remain before reaching commercial smart glasses
The measurements this time used an internal build of the QNN SDK that supports 1-bit computation.
Even if a model's weights can be shrunk, the same speed cannot be achieved unless the device itself has computation functionality available to process them efficiently.
For general developers to try a configuration similar to Bonsai, it needs to become clear which QNN SDK supports 1-bit computation, which devices are compatible, and how a runtime environment including camera input will be provided.
PrismML has not announced provisioning plans reaching that level this time.
It also hasn't been confirmed whether the same configuration can be reproduced as-is on currently available smart glasses on the market.
To compare actual usability, one would need to measure, on the same smart glasses, the time from when the camera captures an image to when the AI returns an answer.
This is because the token generation speed of the language model portion alone cannot determine how long a wait would actually occur in a real visual question-answering scenario.
Toward commercialization, there are points to verify beyond model performance.
Continuously capturing images from the camera consumes power for both image processing and AI inference, and also affects heat generation in the smart glasses.
This material does not include data on battery life, overall device power consumption, or processing speed during extended operation.
If AI processing can be completed entirely on-device, it may become possible to avoid sending images captured by the camera to the cloud every time.
However, it is not guaranteed that voice input and conversation history will also be processed entirely on-device in an actual product. What information is processed on-device and what is sent to the cloud is determined by the design of each individual product that adopts this technology.
In June 2026, Qualcomm announced "Snapdragon START," a development platform for smart glasses makers.
In addition to semiconductor modules, it also provides a software foundation that integrates with smartphone apps and cloud services, aiming to shorten the time it takes for manufacturers to develop products.
However, this is not an announcement that Bonsai has been adopted in a product using Snapdragon START. A configuration combining on-device AI with cloud AI also remains one of the options going forward.
This demonstration by PrismML showed that a roughly 2-billion-parameter AI combining images and language can run on the existing Snapdragon AR1 Gen 1.
What we want to confirm next is whether a development environment supporting 1-bit computation will be made available to general developers, how long it actually takes from capturing an image to receiving a response, how much battery consumption and heat generation occur during extended use, and whether a product actually featuring Bonsai will emerge.
Once those points become clear, it will be possible to judge how far the feature of "AI answering on the spot about what the glasses are looking at" has progressed from a technical demo toward something usable in everyday life.