PrismML, a pioneer in high-performance AI models, today announced Bonsai 27B, PrismML’s largest and most capable Bonsai release to date, available in both 1-bit binary and 1.58-bit ternary variants.
Bonsai 27B, based on Qwen3.6 27B, brings PrismML’s end-to-end low-bit architecture to a new class of model: a multimodal model – accepting images as well as text – built for serious reasoning, coding, and agentic workflows, while compressed small enough to run where conventional 27B-class models would be impractical.
This is the central breakthrough: a model class that previously belonged in the cloud can now move onto the device. At 3.9GB memory footprint, the 1-bit variant of Bonsai 27B is designed to make advanced local intelligence possible on consumer hardware like mobile and laptops.
Bonsai 27B is not a lightweight chat model, a demo, or a proof of compression in isolation. It is built to do real work: reasoning through complex tasks, planning multi-step workflows, writing and debugging code, using tools, and supporting agentic execution locally.
A New 27B Model Series
The Bonsai 27B series is built on PrismML’s Intelligence Density philosophy: delivering the most capability possible per unit of size, memory, power, and deployment footprint. For years, the assumption has been that more capable models require larger infrastructure: more memory, more compute, more power, and more dependence on cloud inference. Bonsai 27B challenges that assumption directly. It shows that the next innovation in AI is not only bigger models in larger data centers, but also more capable models that can run privately, locally, and continuously on the devices people already use.
The series includes two variants:
1-bit Bonsai 27B is optimized for maximum compression and local deployment. At about 4GB, it is designed to make 27B-class intelligence small enough to run locally on high-end mobile devices, such as the iPhone 17 Pro. In fact it achieves 11 tokens/s on iPhone 17 Pro. This represents the most aggressive deployment point in the series: advanced reasoning, vision, and agentic capability moving from the cloud directly onto the device.
Ternary Bonsai 27B uses PrismML’s 1.58-bit ternary weight representation, where each weight is constrained to one of three values: {-1, 0, +1}. It is optimized for higher quality while remaining dramatically smaller (9x) than full precision, and achieves more than 95% of full-precision benchmark performance. The 1-bit variant achieves more than 90%, giving users a practical choice between maximum portability and maximum capability.
Together, the two variants make Bonsai 27B a flexible model series for local intelligence: compact enough for local deployment, capable enough for serious engineering work, and efficient enough to make 27B-class models practical outside traditional cloud infrastructure.
Built for Thinking, Coding, Vision, and Agentic Workflows
Bonsai 27B ships with thinking enabled, reflecting how advanced models are now used in practice: for tool use, planning, coding, and agentic execution.
That distinction matters. Local models have often been framed as lightweight assistants: useful for simple chat, constrained tasks, or offline demos. Bonsai 27B is aimed at a different category. It is designed to be a local workhorse model for software development, debugging, planning, refactoring, tool use, and autonomous workflows. And because Bonsai 27B is multimodal, those workflows are not limited to text: the model can read screenshots, documents, and camera input directly on the device.
A 27B-class model that fits on a phone changes what “local AI” can mean. It makes it possible to imagine advanced assistants that are private by default, available without a network round trip, and capable of running continuously on personal devices. For developers and power users, it opens the door to serious local workflows that do not depend entirely on cloud inference.
Extending the Bonsai Frontier
1-bit and Ternary Bonsai 8B demonstrated that true end-to-end low-bit models could operate across embeddings, attention layers, MLP layers, and the LM head without higher-precision escape hatches. Bonsai 27B extends those ideas to a larger and more capable scale. The binary variant pushes portability to the limit. The ternary variant offers a higher-quality point on the same Pareto curve.
Both are part of a single story: frontier-class capability can become an order-of-magnitude smaller, more efficient, and deployable.
Technical Details:
The 1-bit and Ternary Bonsai 27B models are 27.8-billion parameter multimodal language models. They have been trained using Google v5 TPUs. They have been designed for seamless integration with existing AI workflows and are optimized for low-latency inference on consumer-grade CPUs, and edge GPUs.
Pricing and Availability:
Developers, researchers, and other users can download the Bonsai 27B models under the Apache 2.0 license for free starting July 14, 2026.
About PrismML:
PrismML is a U.S.-based artificial intelligence company focused on making AI more efficient and accessible. PrismML is built on proprietary Caltech intellectual property and backed by Khosla Ventures, Cerberus Ventures, and compute grants from Google and Caltech. For more information, visit the Website, LinkedIn, or X.
####
All registered trademarks and product identifiers belong to their respective corporate entities. Any other trademarks or product names referenced here are also owned exclusively by their relevant companies.