California startup PrismML has compressed a language model with 27 billion parameters, Bonsai 27B, so effectively that it runs on an iPhone. The smallest variant occupies just 3.9 gigabytes of storage, while the uncompressed original requires about 54 gigabytes. PrismML released the model on July 14, 2026, under the Apache 2.0 open-source license.
Two compression tiers cut storage by up to 93 percent
PrismML builds Bonsai 27B on Alibaba’s open base model Qwen3.6-27B and reduces its weights through quantization to one to 1.58 bits instead of the usual 16 bits. The ternary variant, with values of minus one, zero, and plus one, comes to 1.71 effective bits per weight and takes up 5.9 gigabytes. The more aggressive one-bit variant, with pure plus-one and minus-one values, needs only 1.125 effective bits and fits into 3.9 gigabytes on an iPhone 15 or newer. Group-wise scaling at 16-bit precision across embeddings, attention layers, and the output layer is meant to preserve accuracy despite the heavy compression. According to PrismML, the model keeps a 262,000-token context window, processes images through a compact vision encoder, and supports agentic capabilities such as tool calls and multi-step reasoning. On an iPhone 17 Pro Max, the company says the model generates around eleven tokens per second at about 672 tokens per percent of battery charge. On an Nvidia RTX 5090 graphics card, it reaches up to 163 tokens per second. These performance figures come from PrismML itself and are independently unverified.
Apple is testing the technology for its own devices
PrismML founder Babak Hassibi confirmed, according to a report by TechTimes, that Apple and other unnamed companies are testing his startup’s models. The talks are at an early stage but progressing well, he said. Apple is reportedly evaluating the speed, energy efficiency, and on-device performance of the compressed models. The context is the heavy reliance of voice assistants on cloud computing: Bank of America estimates that cloud AI accounts for up to 67 percent of Siri’s weighted computing costs even though it handles only about five percent of queries. A model running directly on the device could answer most of those queries without contacting a server, cutting both latency and operating costs. PrismML was founded by former Caltech professor Hassibi and raised $16.25 million in seed funding from Khosla Ventures, Cerberus Capital, and Caltech before the release. For a startup this size, a trial run with a company like Apple counts as an unusually early win.
Developers get immediate access via GitHub and Hugging Face
Bonsai 27B is available for download right away through Hugging Face and a public GitHub repository. PrismML is also offering a free, limited-time developer preview through an API hosted by Together.ai, letting developers test the model without their own hardware. Besides iPhones, the company says the model also supports laptops via the llama.cpp and MLX frameworks, as well as Nvidia GPUs via CUDA kernels built for low bit widths. Interested users can try the model directly in a browser through a WebGPU demo on Hugging Face Spaces, without setting up any hardware. An iOS app called Locally AI already integrates Bonsai 27B. The open Apache 2.0 license permits commercial use and custom modifications without licensing fees, which should help the model spread well beyond PrismML’s own channels.
What matters now is whether the performance figures PrismML cites hold up under independent testing, and whether Apple actually builds the technology into a shipping product. If on-device compression at this scale becomes standard, it would pressure the business model of cloud-based AI providers, whose revenue has so far depended heavily on data center capacity.
