#United States #Pasadena #PrismML #Bonsai2 #AIModels

PrismML Enhances Bonsai 2 Lineup with New Models

On October 15, 2026, PrismML made headlines in the AI industry by expanding its Bonsai 2 family with the introduction of three new models: 2B, 4B, and 9B. These advanced ternary models are built upon the next-generation Qwen3.5 architecture, promising enhanced performance and efficiency tailored for the resource limitations found in everyday devices like smartphones, personal computers, and wearables.

A Leap in Technology

The latest offerings from PrismML emphasize powerful capabilities without taxing the memory and power resources of devices. By continuing the legacy of the first-generation Bonsai models, which included configurations of 1.7B, 4B, and 8B parameters, the new 2B, 4B, and 9B models now encompass a complete range of sizes from 2B up to a robust 27B.

PrismML adopts a unique ternary weight representation for these models, employing the values {-1, 0, +1} to encode model weights. This leads to a significantly reduced model footprint, where the new models weigh in at 0.43GB, 0.95GB, and 2.00GB respectively—over nine times smaller than their full-precision counterparts.

Babak Hassibi, Founder and CEO of PrismML, highlights the importance of creating on-device agents that can operate effectively while adhering to strict memory and power constraints. He states, "With these Ternary Bonsai 2 releases, we're giving developers more capability within those constraints and more flexibility in what they can build."

Superior Performance for Localized AI

In rigorous testing across 19 benchmarks, the new Ternary Bonsai 2 models have either matched or exceeded the performance of traditional full-precision models, particularly in tasks related to coding, mathematics, and instruction adherence. For scenarios that require visual understanding and reasoning, these models have impressively retained over 90% of the benchmark scores when pitted against their full-precision Qwen3.5 counterparts. Enhanced inference speeds also empower quicker decision-making abilities.

With powerful hardware like the RTX 4090 and the Macbook M5 Pro, the Bonsai 2 models demonstrate up to three times the energy efficiency and five times the throughput compared to equivalent full-precision models. The 9B model can achieve up to 124 tokens per second (toks/s), while the 2B model reaches an impressive 276 toks/s.

Furthermore, these compact models are well-suited for deployment on mobile devices. For instance, the Bonsai 2 9B performs at up to 34 toks/s on a cutting-edge iPhone 18 Pro Max, while the 2B version performs at a robust 120 toks/s.

Building Next-Generation Applications

At the Snapdragon Summit last September, PrismML showcased an innovative 1-bit Bonsai vision-language model operating efficiently on smart glasses powered by the Snapdragon AR1 Gen 1 Platform. This demonstration illustrated the potential for visual understanding and assistance within wearable hardware while remaining mindful of the constraints typical of such devices.

The new Bonsai 2 models further enhance these capabilities, supporting a wider range of applications that leverage improved reasoning, visual interpretation, and tool utilization. These models pave the way for developing advanced digital assistants capable of intricate user interaction, allowing them to interpret requests, reason through context, and execute a series of actions using shared digital tools.

Technical Blueprint

The Ternary Bonsai 2 models—comprising 2B, 4B, and 9B—have been expertly trained utilizing Google v5 TPUs. They are optimized for seamless integration with existing AI workflows and designed for low-latency inference even on consumer-grade CPUs and edge GPUs.

Access and Pricing

Starting from October 15, 2026, developers and researchers can freely access the Bonsai 2 models under the Apache 2.0 license, allowing them to explore and implement cutting-edge AI capabilities in their projects.

For further engagement, users can download the Ternary Bonsai 2 models from PrismML's website.

About PrismML

PrismML stands as a pioneering U.S.-based AI company focused on making artificial intelligence more efficient and accessible to all users. Emerging from a collaboration of researchers from Caltech, the company enjoys support from notable investors including Khosla Ventures, Cerberus, and Google, along with ongoing backing from Samsung. Learn more about PrismML through their official channels including their website, LinkedIn, and X.

• ---

This latest development marks a significant milestone in the field of artificial intelligence, reflecting PrismML's commitment to innovation and accessibility in tech solutions.