The race to bring powerful artificial intelligence directly onto consumer devices has taken a major step forward. California-based AI startup PrismML has introduced Bonsai 27B, an open-source multimodal language model designed to run on smartphones, potentially reshaping how advanced AI applications operate without relying heavily on cloud infrastructure.
The launch comes amid reports that Apple is evaluating PrismML’s model-compression technology as it explores ways to run larger AI models directly on future iPhone devices, highlighting growing industry interest in efficient on-device AI.
Bringing Large-Scale AI to Smartphones
Traditionally, large language models with billions of parameters require significant computing power and memory, making them impractical for smartphones and many personal computers. PrismML claims its newly released Bonsai 27B overcomes this limitation through advanced compression techniques that dramatically reduce storage and memory requirements.
Built on Alibaba’s Qwen3.6 27B foundation model, Bonsai 27B is available in two optimized versions:
- A ternary model requiring approximately 5.9 GB of storage
- A 1-bit version occupying just 3.9 GB
For comparison, a standard 27-billion-parameter model typically demands around 54 GB of memory when running in FP16 precision, while even a compressed 4-bit version generally requires about 18 GB. By shrinking the model footprint to smartphone-friendly levels, PrismML aims to make advanced AI capabilities accessible on everyday devices.
Advanced Capabilities Without the Cloud
PrismML positions Bonsai 27B as the flagship model in its Bonsai family, offering features commonly associated with much larger cloud-based AI systems.
The model supports:
- Multimodal inputs
- Multi-step reasoning
- Structured tool usage
- Long-context workflows
- Agentic task execution
- Speculative decoding
- End-to-end low-bit processing
One of the model’s standout features is its 262,000-token context window, allowing it to handle significantly larger amounts of information during interactions.
According to the company, these capabilities make the model suitable for complex applications including software development, reasoning tasks, automation workflows, and AI agents operating locally on devices.
Performance Remains Strong Despite Compression
A major challenge in model compression is preserving performance while reducing size. PrismML says its compression techniques achieve this balance effectively.
The company reported that:
- The ternary version maintains approximately 95% of the original model’s performance
- The 1-bit variant retains roughly 90% performance
These figures were measured across a suite of 15 benchmark evaluations, suggesting that users can access advanced AI capabilities without substantial losses in accuracy or functionality.
Apple’s Interest Signals Growing Demand for On-Device AI
The launch gains additional significance as reports indicate that Apple has been assessing PrismML’s compression technology for future AI initiatives.
As smartphone manufacturers increasingly focus on privacy, efficiency, and offline AI experiences, running sophisticated models directly on devices has become a strategic priority. Local processing can reduce cloud costs, improve response times, enhance data privacy, and enable AI features even when internet connectivity is limited.
If adopted, technologies like Bonsai 27B could help bridge the gap between powerful cloud-based AI systems and mobile hardware constraints.
Open-Source Approach Could Accelerate Adoption
Both Bonsai 27B variants are being released under the Apache 2.0 open-source license, allowing developers and organizations to experiment with, modify, and deploy the models across a wide range of applications.
PrismML believes the industry has long lacked a model that combines compact size with sufficient intelligence for demanding workloads.
The company argues that while cloud-based AI services will continue to play an important role, organizations increasingly need capable local models that can operate independently and efficiently on consumer hardware.
A Fast-Rising AI Startup
PrismML was founded in 2026 as a spinout from the California Institute of Technology (Caltech) by Babak Hassibi, Sahin Lale, Omead Pooladzandi, and Reza Sadri. Despite being a relatively new entrant in the AI sector, the startup has already attracted significant investor attention.
The company has secured $16.25 million in seed funding from investors including Khosla Ventures, Cerberus Ventures, and Caltech, while also receiving backing from Google and ongoing support from Samsung.
Earlier this year, PrismML introduced its Bonsai 8B family, demonstrating similar low-bit compression techniques for smaller language models. With the debut of Bonsai 27B, the company is now targeting a much larger class of AI models, potentially setting a new benchmark for smartphone-ready artificial intelligence.
As demand for on-device AI continues to accelerate, PrismML’s latest breakthrough could play a key role in bringing enterprise-grade AI capabilities directly into consumers’ pockets.
Read more:Â She Built It Herself: The Making of Nere De Achurra
