The smarter AIs get, the larger they become. They also require more computing power and storage space to run properly.
This is why powerful large language models (LLMs) typically run on massive clusters of GPUs inside data centers, rather than on laptops or smartphones. Every new generation tends to increase in parameter count, memory requirements, and computational demands, making training and inference increasingly expensive.
But a new wave of research is challenging that assumption.
Instead of relying on ever-larger models, developers are finding ways to make AI smaller, more efficient, and capable of running directly on personal devices. By compressing models, optimizing architectures, and using specialized hardware, AI can increasingly operate offline, respond faster, consume less energy, and better protect user privacy, all without always needing a connection to the cloud.
And PrismML aims for that.
PrismML is a U.S.-based AI startup that emerged from stealth in March 2026 with a very specific mission: make frontier AI models dramatically smaller and more efficient, without sacrificing capability.
Instead of competing to build the LLMs, PrismML focuses on what it calls "intelligence density," or maximizing reasoning ability per unit of memory, compute, and energy.
And 'Bonsai 27B' came out of that idea.
According to the announcement, the model is literally a pair of compressed versions of Alibaba's open-source Qwen 3.6 27B model that bring a new scale of capability to devices that previously could not host anything close to this size.
The first one, called the Ternary Bonsai 27B, uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, giving a true 1.71 effective bits per weight. The second one, the 1-bit Bonsai 27B, uses binary {−1, +1} weights with the same group-wise scaling, giving 1.125 effective bits per weight.
The original model occupies roughly 54 GB when stored in 16-bit precision.
Even aggressive conventional 4-bit quantization leaves it around 18 GB, still too large for phones and for many laptops.
Bonsai 27B reduces that footprint to 5.9 GB in its ternary form and 3.9 GB in its 1-bit form, making the latter the first 27B-class model able to fit inside the memory budget of a modern iPhone.
"On phones, the fitting constraint is tighter than raw RAM suggests. iOS limits any single app to roughly half of physical memory, so a 12 GB iPhone exposes only about 6 GB to the model. That rules out FP16 and Q4_K_XL, and even the ternary build (5.9 GB) would nearly the entire budget with nothing left for the KV cache and activations - leaving 1-bit Bonsai at 3.9 GB as the only variant that runs on-device with headroom," the whitepaper says.
As LLMs have grown more capable, their parameter counts and memory requirements have climbed in parallel.
Stronger reasoning, broader knowledge, longer context windows, and the ability to call tools or process images all add parameters. Those parameters are normally stored with enough numerical precision that the resulting files demand server-class GPUs or large amounts of RAM.
Running a full 27B model locally has therefore remained impractical for most users; the compute and storage costs simply do not fit consumer hardware.
Bonsai 27B attacks that constraint by representing each weight with only one or three possible values (binary or ternary) plus a small amount of scaling information.
The entire network, including embeddings, attention layers, feed-forward blocks, and the language-model head, stays in this low-bit format. No high-precision islands remain to inflate the memory footprint.
The resulting models retain most of the original's measured ability.
Across a 15-benchmark suite that covers knowledge, math, coding, instruction following, tool use, and vision, the ternary variant scores about 95% of the full-precision baseline while the 1-bit variant scores about 90%. Math and coding performance stay especially close to the original; tool-calling and multi-step agentic behavior also remain usable.
Both versions support a 262 000-token context window, multimodal inputs through a compact vision tower, structured tool calls, and coherent multi-step agent loops.
They can therefore handle the kinds of sustained workflows that now dominate practical AI use: reading local files, calling external tools, updating internal state, and iterating without constant round-trips to a remote server.
Because the models fit on ordinary hardware, those workflows can stay entirely local.
Private documents never leave the device, intermediate reasoning states remain offline, and the marginal cost of additional steps drops to zero once the model is loaded.
The 1-bit version is small enough to share an iPhone's limited application memory with its key-value cache, activations, and the rest of the operating system.
The ternary version targets laptop-class machines where higher quality is preferred.
Both run natively through Apple's MLX framework on Macs, iPhones, and iPads, and through custom CUDA kernels on Nvidia's GPUs.
Apple has taken notice.
According to reporting published the same day as the release, the company is in early-stage discussions with PrismML to evaluate the compression technology for on-device use.
For Apple, which already routes many simple AI tasks to the Neural Engine and keeps more sensitive work in its private cloud, a reliable way to run a 27B-class multimodal model inside an iPhone would expand what Siri and other system features can do without leaving the device.
Lower latency, stronger privacy guarantees, and reduced dependence on remote compute are the practical consequences.
The release is fully open under the Apache 2.0 license, with model weights, inference code, and a technical white paper available immediately.
