The launch has drawn attention after a CNBC report said Apple has been in talks with PrismML to evaluate its model-compression technology, which aims to shrink large language models so they can run natively on iPhones.
Neither company has confirmed a partnership, but the discussions highlight growing interest in running more capable AI systems directly on devices.
What is Bonsai 27B
According to PrismML, Bonsai 27B is not a new foundation model. It is a compressed implementation of China-based Alibaba’s Qwen3.6 27B model. The company says it retained the original architecture while significantly reducing the model’s memory footprint, allowing it to run on consumer hardware that could not previously accommodate a model of this size.
Rather than pre-training a new model, PrismML focused on making an existing large model considerably smaller without changing its core architecture.
For context, Qwen3.6 27B is Alibaba’s open-weight, 27-billion-parameter multimodal AI model, designed for coding, reasoning and multimodal tasks while remaining easier to deploy than larger models.
PrismML says Bonsai 27B introduces a new capability tier for local AI.
According to the company, it supports multi-step reasoning, structured tool calling, long-context workflows and coherent agentic loops — capabilities generally associated with much larger, cloud-based AI systems.
Why large AI models struggle to run on phones
According to PrismML, the main obstacle is model size. A conventional 27-billion-parameter model requires about 54 GB of storage in 16-bit precision, making it impractical for smartphones and difficult to deploy on many laptops.
The company says that even after conventional 4-bit quantisation, the model still occupies about 18 GB, exceeding the storage and memory limits of most consumer devices.
These hardware constraints have largely confined models in the 27-billion-parameter capability class to cloud infrastructure or specialised computing systems.
PrismML says Bonsai 27B is designed to overcome this problem by sharply reducing the model’s footprint while retaining much of its original capability.
Bonsai 27B comes in two versions
PrismML has released Bonsai 27B in two variants, each designed for a different category of device. The Ternary Bonsai 27B occupies 5.9 GB and uses 1.71 effective bits per weight.
According to PrismML, it is optimised for what the company calls laptop-class quality, making it suitable for notebooks and other devices with comparatively greater computing resources.
The second version, 1-bit Bonsai 27B, occupies 3.9 GB and uses 1.125 effective bits per weight. The company says this version is optimised for phone-class deployment, making it small enough to run on modern smartphones while retaining much of the capability associated with a 27-billion-parameter model.
How quantisation makes AI models smaller
Quantisation is a technique used to make AI models smaller and more efficient without rebuilding them from scratch. In simple terms, it reduces the memory required to store an AI model by representing its numerical values, known as weights, using fewer bits. Because each weight occupies less space, the overall model becomes smaller and can run on devices with limited memory and computing power, including smartphones and laptops.
PrismML says Bonsai 27B uses extremely low-bit representations. The Ternary Bonsai 27B uses three possible weight values — minus one, zero and plus one — resulting in 1.71 effective bits per weight.
The 1-bit Bonsai 27B uses two possible values — minus one and plus one — reducing the requirement to 1.125 effective bits per weight.
According to the company, this process reduces the model’s size from about 54 GB in conventional 16-bit precision to 5.9 GB for the laptop version and 3.9 GB for the smartphone version.
That makes it possible, PrismML says, to run a 27-billion-parameter-class AI model directly on consumer devices.
What PrismML claims Bonsai 27B can do
PrismML says Bonsai 27B is designed to improve what it calls “intelligence density”, or the amount of AI capability that can be delivered within a given amount of memory. The company argues that while raw capability determines what a model can do, intelligence density determines where it can run.
PrismML claims Bonsai 27B delivers 27-billion-parameter-class capability in a footprint smaller than many full-precision two-billion-parameter models, making advanced AI more practical on consumer devices rather than limiting it to servers.
The company also says both Bonsai 27B variants are multimodal, allowing them to process screenshots, documents and camera inputs, in addition to text, through a compact 4-bit vision tower.
According to PrismML, the models support a 262,000-token context window, speculative decoding for faster inference and run entirely in their low-bit representation across the language network without reverting to higher-precision components.
The company says these capabilities are intended to enable local AI workflows on smartphones and laptops.
Bonsai 27B is available as open-source software under the Apache 2.0 licence.
Apple link
According to CNBC, Apple has held discussions with PrismML to evaluate its AI model-compression technology. The technology is designed to shrink LLMs so that they can operate within the memory and storage limits of devices such as the iPhone.
PrismML’s approach aims to reduce model size significantly without relying entirely on cloud processing. If successful, such technology could allow larger AI models to run natively on consumer hardware, supporting Apple’s broader effort to expand on-device AI capabilities while reducing dependence on remote servers.
Neither Apple nor PrismML has announced a partnership or acquisition.
