New Delhi: PrismML has launched Bonsai 27B, a compressed artificial intelligence model that the company says can run locally on smartphones and laptops. The model comes in 1-bit binary and 1.58-bit ternary versions, with the smaller version using about 3.9GB of memory.
The model is based on Qwen3.6 27B and can process both text and images. PrismML says it can handle coding, reasoning, screenshots, documents and tool-based tasks without sending every request to a cloud server. The 1-bit version reportedly generated around 11 tokens per second on an iPhone 17 Pro.
A normal 27-billion parameter AI model can need several times more memory, depending on the format used. PrismML cuts the size by storing model weights using only one bit for its most compressed version.
Think of model weights as billions of tiny settings that help an AI decide what answer comes next. Most models store these settings with far more numerical detail. Bonsai reduces that detail heavily, which lowers storage and memory use.
The ternary version uses three possible values for each weight: -1, 0 and +1. It takes more space than the binary version but aims to preserve more of the original model’s performance.
PrismML describes this approach as delivering the “most capability possible per unit of size, memory, power, and deployment footprint.”
Source: PrismML
Across 15 tests covering maths, coding, knowledge, instruction following, tool use and vision, the full-precision Qwen model scored an overall 85.
The results shared by PrismML were:
Math performance stayed relatively close. The full model scored 95.3, compared with 93.4 for the ternary model and 91.7 for the 1-bit model.
The bigger drop came in tool use and vision. The 1-bit version scored 66 in agentic and tool-calling tests, compared with 80 for the full model. In vision tests, it scored 59.6 against 72.6.
These figures suggest compression still carries a quality cost, especially for tasks that need image understanding or external tools.
Running an AI model on a phone can reduce dependence on an internet connection. It may also keep screenshots, documents and personal data on the device.
This could help developers build offline coding assistants, private document readers or camera-based tools. Real-world performance will depend on battery use, heat, memory limits and software support.
PrismML says Bonsai 27B was trained using Google v5 TPUs and supports consumer CPUs and edge GPUs. The models were released under the Apache 2.0 licence on July 14, 2026.
The launch follows PrismML’s earlier Bonsai 8B model, which compressed an 8.19-billion parameter model from 16.38GB in FP16 format to about 1.15GB in GGUF format. That model showed the same basic trade-off. Much smaller files, faster local processing, and some loss in benchmark scores.
Click for more latest Artificial Intelligence news. Also get top headlines and latest news from India and around the world at News9.
Siddharth Shankar is a journalist with over a decade of experience covering business, infrastructure, technology, science, gaming, automobiles and digital media. Currently leading the newsroom, with a focus on Business, Technology, AI, Cybersecurity, Science, Gaming and Automobile sections. For news leads, he can be reached on X at @Siddh4rth.