A 27.8-billion-parameter AI model now fits on your phone.

PrismML, a Caltech-spinout AI company backed by Khosla Ventures, Cerberus, Google, and Samsung, announced Bonsai 27B on July 14. The model runs locally on high-end mobile devices like the iPhone 17 Pro, processing roughly 11 tokens per second in its most compressed form. No cloud. No latency round-trips. Just a phone doing the heavy lifting.

How you shrink a giant model into a pocket

The trick is something called extreme low-bit quantization. PrismML reduced the precision of each parameter from 16-bit floating point down to as low as 1 bit.

A standard FP16 version of the model would occupy around 54 GB of memory. The 1-bit binary version of Bonsai 27B takes up just 3.9 GB. The 1.58-bit ternary variant sits at 5.9 GB. Despite that aggressive compression, PrismML says the model retains 90-95% of benchmark performance compared to its full-precision baseline across math, coding, reasoning, and vision tasks.

Bonsai 27B is built on top of the Qwen3.6 27B architecture and was trained on Google v5 TPUs. It handles multimodal inputs, meaning it processes both text and images. It supports a 262,000-token context window. The model also includes tool-calling capabilities for agentic workflows, the kind of setup where an AI can autonomously browse the web, pull data, or execute multi-step tasks.

The road to 27 billion parameters on a phone

PrismML has been scaling up its Bonsai model family throughout 2026. The company released an 8B parameter model in March, followed by Image 4B in May. Each release served as a proving ground for the quantization techniques that now power the 27B version.

The model ships under the Apache 2.0 open-source license, which means developers can download it for free and build on top of it without restrictive terms. PrismML is also offering a limited developer preview API for those who want to test capabilities before committing to local deployment.

Why crypto and fintech investors should pay attention

On-device AI eliminates one of the biggest friction points in crypto: the trust problem with cloud infrastructure. When AI models run locally, user data never leaves the device. For crypto wallets, DeFi interfaces, and trading tools that increasingly rely on AI-powered analysis, that’s a meaningful privacy upgrade.

The agentic workflow capabilities are equally relevant. A 27B-parameter model running locally, with tool-calling abilities and a 262K context window, is exactly the kind of foundation autonomous agents need for executing trades, managing yield farming positions, and monitoring on-chain activity.

Apple, Google, and Qualcomm have all been pushing on-device AI, but mostly with smaller models in the 3-7B parameter range. PrismML just leapfrogged that entire class by nearly 4x in parameter count while keeping the memory footprint manageable for flagship phones.

The open-source Apache 2.0 license is a key factor here. Unlike proprietary models locked behind API paywalls, Bonsai 27B can be forked, fine-tuned, and embedded into any application. Crypto developers building wallet assistants, on-chain analytics tools, or autonomous trading agents now have a foundation model they can run without cloud costs, without data leakage, and without asking permission.