PrismML expands its Bonsai 2 family

PrismML releases three Bonsai 2 models with 2, 4, and 9 billion parameters, designed to run AI agents on smartphones and computers.

A nine-billion-parameter model occupying just 2 GB, capable of interpreting images, reasoning, and using tools directly from an iPhone. With Bonsai 2, PrismML expands its line of compact models with three new versions featuring 2, 4, and 9 billion parameters. Derived from Qwen3.5, they use a ternary weight representation to reduce their memory requirements. The company is targeting local assistants capable of performing operations on a smartphone, computer, or connected device without systematically transmitting their data to a remote server. Three New Models in Under 2 GB

The new line includes Bonsai 2 2B, 4B, and 9B, joining a 27-billion-parameter version previously introduced by PrismML.

These models succeed the first-generation Bonsai 1.7B, 4B, and 8B, with announced improvements in reasoning, visual understanding, and tool use.

Their distinctive feature lies in a weight representation using only three values: −1, 0, and +1.

This technique, known as ternary quantization, drastically reduces the amount of memory required to store parameters compared to traditional full-precision numerical representations.

PrismML announces sizes of 0.43 GB for Bonsai 2 2B, 0.95 GB for the 4B, and 2 GB for the 9B. These footprints are more than nine times smaller than those of the corresponding full-precision Qwen3.5 models.

The memory reduction facilitates their loading on resource-constrained devices, though these figures do not necessarily represent total runtime consumption. Context, intermediate data, and the software environment also consume memory. Assistants Capable of Acting Directly On-Device

PrismML focused this generation on several behaviors designed for personal assistants: following instructions, avoiding repetitive answers, limiting unnecessary loops, and calling tools available within an application.

The goal is to allow models to engage in tasks requiring multiple operations, rather than being limited to answering questions.

On a phone, an assistant could analyze a document, extract specific information, and then pass it to a local application. Another use case involves interpreting an image and using the features available on the device to carry out a task.

PrismML also worked on response conciseness. According to its evaluations, Bonsai 2 9B uses approximately 40% fewer tokens on common instruction-following and math tasks.

This reduction applies to the scenarios measured by the company and does not guarantee identical consumption across all applications. Up to 120 Tokens Per Second on iPhone

The models' reduced size enables them to run on recent mobile devices.

Specifically, PrismML showcases a demonstration of Bonsai 2 9B running locally on an iPhone 18 Pro Max, complete with visual interactions and tool calls.

On this device, the 9B model achieves up to 34 tokens per second, while the 2B version reaches up to 120 tokens per second.

Announced performance is also high on Apple computers. On a configuration designated M5 Max Pro by PrismML, Bonsai 2 9B reaches up to 179 tokens per second, compared to 310 tokens per second for Bonsai 2 2B.

The company also reports measurements on configurations equipped with NVIDIA RTX 4090s and Apple M5 Pro processors. According to its tests, the ternary models can reach up to five times the throughput and three times the energy efficiency of their full-precision Qwen3.5 counterparts.

These results correspond to specific hardware and execution conditions. Performance varies based on factors such as model size, context, and available resources.

For mobile applications, energy efficiency is as important as speed. A model that consumes fewer resources can be queried more frequently without imposing the same drain on the battery. Nineteen Evaluations Against Qwen3.5 Models

PrismML compared its three new models to their Qwen3.5 counterparts across 19 evaluations spanning six domains: tool use, mathematics, instruction following, coding, knowledge and reasoning, and visual understanding.

The results show that compression does not affect all capabilities equally.

On BFCLv3, dedicated to tool calling, Bonsai 2 9B scores 71.81 compared to 73.81 for Qwen3.5 9B. On MATH-500, scores reach 98.60 and 99.00, respectively.

In other domains, the differences are more pronounced. On GPQA Diamond, Bonsai 2 9B reaches 69.9 compared to 82.53 for its reference model.

PrismML also publishes performance retention rates by domain, established by comparing the scores of the ternary models to those of the original versions.

In visual understanding, the reported overall rates are 89% for Bonsai 2 2B, 91.3% for the 4B, and 92.5% for the 9B.

Instruction following shows different results: 121.5% for the 2B, 106% for the 4B, and 100% for the 9B. On the selected evaluations, certain compressed versions thus match or even surpass their full-precision baselines.

Coding also retains close results, with rates of 115%, 100.7%, and 100.1% depending on model size.

Knowledge and reasoning are impacted more significantly: 82.9% for the 2B, 90.9% for the 4B, and 93% for the 9B.

These gaps show that memory reduction does not uniformly preserve different capabilities. Tasks involving specialized knowledge or complex reasoning can be more sensitive to compression.

The results stem from the evaluations published by PrismML and do not constitute a general measurement of performance in real-world applications. Local Processing Before Offloading to the Cloud

PrismML also presents a demonstration combining local artificial intelligence with a remotely hosted model.

In this example, Bonsai 2 9B examines confidential documents on a Mac. The model redacts certain sensitive information with codes before another AI, running in the cloud, continues the analysis and drafts a due diligence memo.

According to PrismML, only 8.6% of the initial data is transmitted to the cloud in this setup, with a reported 86% reduction in costs associated with remote processing.

This approach makes it possible to reserve certain operations for the local