PrismML Squeezes a 27B Model to Fit on a Laptop
PrismML raised a $22.25 million seed round and released Bonsai 2 27B, a compressed version of Alibaba's widely used open-source Qwen3.8 27B model. The announcement was reported on Sept. 17, 2026. TechCrunch
Bonsai 2 27B compresses Qwen3.8 27B down to 5.9 GB. That is a 9x to 10x reduction in memory versus the original model. PrismML reports it matches 98% of Qwen's aggregate benchmark scores.
The compression uses ternary weights of +1, -1, or 0. Standard models store each weight at 16 bits. Ternary quantization cuts storage per parameter sharply while preserving matrix structure for inference, the step when the model generates answers. The result fits in laptop memory. It avoids datacenter-class GPUs for serving.
PrismML was founded by Caltech researchers and is led by CEO Babak Hassibi, a Caltech professor. Ion Stoica serves as an advisor. The company is backed by Khosla Ventures, Cerberus Capital, and Caltech.
This is the company's second compression step. Its first Bonsai model was released in March and matched 95% of benchmark scores. That model has been downloaded over 11 million times. Even smaller models from the company have been downloaded another 2.6 million times.
PrismML announced its 1-bit Bonsai models on March 31, 2026, describing them as the first commercially viable 1-bit LLMs. Contemporary coverage of that release noted a full-power model fit into 1.15 gigabytes of memory, approximately 14 times smaller than a standard 16-bit model and running 8 times faster.
Bonsai 2 27B is a ternary model. PrismML describes it as its most capable model yet. The company announced Bonsai 27B as its multimodal flagship. It is built to bring agentic coding, multimodal reasoning, and long-horizon tool use to local devices. That means AI that can write code, work with text and images together, and complete multi-step tasks with tools.
Two packaged variants are listed. The ternary version is 5.9GB and runs on laptops. The 1-bit version is 3.9GB and runs on an iPhone 17 Pro. The model has a 262K-token context, the working memory it can use at once. The company states its Bonsai models pack advanced reasoning, tool use, and multimodal AI into phones, laptops, and efficient data centers.
The portfolio extends beyond text. PrismML's Ternary Bonsai is a family of 1.58-bit language models designed to balance strict memory constraints with high accuracy requirements. PrismML released Bonsai Image 4B, a family of compact image-generation models designed to run high-quality diffusion inference, the repeated refining process used to make images. Its site at prismml.com states it builds 1-bit and ternary Bonsai models.
In my view, the pace of improvement is the detail to watch. From 95% in March to 98% now, with a larger base model, the tooling around ternary training and distillation appears to be maturing. Compression work often trades a few points of accuracy for large efficiency gains. Closing that gap makes local agentic workflows more practical, where tool calls and code generation punish small reasoning errors. My kids grew up as speech recognition moved from the lab to on-device dictation that just worked, and capability felt different once it left the server. Local models of this class could follow the same path.
The broader context here is choice of infrastructure. A 27B-class model at 5.9GB changes memory bandwidth requirements, batching behavior, and power draw. It moves capable inference from clustered accelerators to single consumer GPUs and high-end laptops. It also lowers the cost of keeping long contexts resident, since a 262K-token context is expensive at full precision and becomes workable when weights are ternary. Cloud inference will continue for frontier training and heavy multi-tenant serving. Local inference enables offline use, tighter data control, and lower per-token cost at the edge. If developers can rely on 27B-class reasoning without a network round trip, new categories of assistants, coding agents, and multimodal tools become easier to ship.
