PrismML Brings 1-Bit Bonsai Models to AI Smart Glasses Powered by Snapdragon
1-bit Bonsai dramatically increases the on-device intelligence of smart glasses, fitting 4× larger models onto the same memory footprint
PASADENA, Calif. — September 23, 2026 — PrismML today announced a new class of AI capability, specifically designed for smart glasses powered by Qualcomm Technologies’ Snapdragon® platforms.
At the Snapdragon Summit, PrismML’s 1-bit Bonsai model was demonstrated running locally on AI smart glasses powered by the Snapdragon AR1 Gen 1 Platform. Within the same memory constraints, the 1-bit model makes it possible to fit a model with 4× as many parameters as could previously run on some glasses form factors.
The model is a new 2-billion-parameter vision-language model built on PrismML’s Bonsai 1.7B, released earlier this year. Designed for smart glasses, it provides the requisite visual and language abilities to help enable experiences such as understanding what the wearer is seeing, reasoning over that context, and responding in real time.
PrismML worked with Qualcomm Technologies to optimize the model’s weights and architecture specifically for the Qualcomm® Hexagon™ NPU, taking advantage of the underlying hardware architecture to maximize performance and efficiency.
That kind of optimization especially matters on smart glasses. They operate within extremely tight limits on memory, power, and thermal headroom, while still needing to remain responsive and last throughout the day.
“The future of Personal AI depends on bringing powerful intelligence directly to the devices people use every day. By enabling support for PrismML’s innovative 1-bit models on Snapdragon AR1 Gen 1, we’re helping make advanced AI experiences more practical on lightweight, power-optimized wearables,”
said Ziad Asghar, Senior Vice President and General Manager, XR, Wearables and Personal AI, Qualcomm Technologies, Inc.
“This breakthrough opens the door to richer multimodal AI experiences such as vision-based assistance that run locally, with greater responsiveness and privacy-focused features.”
On the Snapdragon AR1 Gen 1 Platform, PrismML’s 1-bit Bonsai delivers the equivalent intelligence of the model at 4-bit precision1, while using roughly 4× less memory2, and generating tokens at more than 2× the speed3. This can make interactions more immediate and natural, while keeping context on the device.
“For AI to become ubiquitous, capable intelligence has to fit into the devices people already carry and wear,”
said Babak Hassibi, Cofounder and CEO of PrismML.
“Smart glasses make that challenge very concrete: you have an extraordinary amount of context available through vision, but an extremely constrained device in which to understand it. Working with Qualcomm Technologies, we designed Bonsai around those constraints. The result is substantially more intelligence within the same hardware envelope. We believe that model-hardware co-design is how AI moves from the cloud into the devices people use every day.”
PrismML and Qualcomm Technologies plan to continue expanding their work together, optimizing increasingly capable Bonsai models across Snapdragon platforms. As AI is moving into glasses, wearables, and the devices around us, the next leap cannot come from adding more compute alone. It requires getting dramatically more intelligence from the compute already there.
About PrismML
PrismML is a U.S.-based artificial intelligence company focused on making AI more efficient and accessible. PrismML emerged from a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus, and Google, with continuing support from Samsung, Gates Frontier, and Jeff Dean. For more information, visit the Website, LinkedIn, or X.
1 It is based on benchmark evaluations conducted by PrismML Inc. in September 2026, comparing the Bonsai 1.7B 1-bit LLM with corresponding Qwen 3 1.7B model with 4bit quantization across BFCL v3, HumanEval+, MMLU Redux, IFEval, IFBench, MuSR, GSM8K, GPQA Diamond, industry-standard evaluation methodology to ensure fair comparisons. The 1-bit and 4-bit configurations achieved comparative benchmark results, respectively, under the specified evaluation conditions. Results may vary depending on model configuration, evaluation methodology, workload, and other applicable conditions.
2 It is based on testing conducted by Qualcomm Technologies International, Ltd. in Sep 2026 on the Snapdragon® AR1 Gen 1 Platform using the Bonsai 2B VLM model, consisting of a 1.7B 1-bit LLM and a 0.3B 4-bit vision encoder, with a context length of 1,024 tokens, compiled using an internal QNN SDK with 1-bit kernel support, on a platform configured with 4 GB memory. LLM weights memory was 0.43 GB for the 1-bit 1.7B LLM model versus 1.66 GB for the corresponding 4-bit 1.7B LLM model, representing a 3.83x reduction (74% smaller). Performance may vary depending on model configuration, software version, hardware configuration, workload characteristics, and operating conditions.
3 It is based on testing conducted by Qualcomm Technologies International, Ltd. in September 2026 on the Snapdragon® AR1 Gen 1 Platform using the Bonsai 2B VLM model, consisting of a 1.7B 1-bit LLM and a 0.3B 4-bit vision encoder, with a context length of 1,024 tokens, compiled using an internal QNN SDK with 1-bit kernel support, on a platform configured with 4 GB memory and a peak AI compute capability of 6 TOPS and 2,304 MACs per cycle. Token-generation rate was 15.36 tokens per second (TPS, low-end number) for the 1-bit 1.7B LLM model versus 7.44 TPS for the corresponding 4-bit 1.7B LLM model, representing a 2.06x improvement. Performance may vary depending on model configuration, software version, hardware configuration, workload characteristics, and operating conditions.
Snapdragon and Qualcomm branded products are products of Qualcomm Technologies, Inc. and/or its subsidiaries. Qualcomm, Snapdragon and Hexagon are trademarks or registered trademarks of Qualcomm Incorporated.
%201.svg)
