A 27-billion-parameter model running entirely on a phone: until yesterday, that would have sounded like science fiction. Today it is reality with Bonsai 27B, the result of work by the PrismML team, who took Qwen3.6-27B and compressed it to just 3.9GB using an extreme binary quantization. The model now takes up less than 4GB, enough to fit into the unified memory of an iPhone 15 Pro Max (8GB) and respond with acceptable latency through the Atomic Chat app.

This is not the usual compression with a few layers kept at higher precision as a safety net. PrismML adopted a true binary scheme called “binary g128”: every weight is reduced to a single sign bit, and a FP16 scale factor is shared among groups of 128 weights, landing at around 1.125 bits per weight. The biggest surprise is that not only the linear layer weights, but also the embeddings, attention and MLP projections, and even the LM head are all quantized to 1 bit with no exceptions. In current practice, most 1-bit schemes preserve some components at higher precision precisely to avoid collapse in the early or final stages of the pipeline. Here, they went all the way.

Benchmark scores across 15 tests average 76.1 compared to the FP16 model’s 85.1 — roughly 89.5% retention. The area that suffers most is knowledge and reasoning, dropping from 83.2 to 73.4: this is where a user might notice missing details or less precise answers. Math, on the other hand, holds up at 91.7, suggesting that procedural and symbolic tasks tolerate extreme weight reduction better.

Memory-wise, the model behaves predictably: around 5.2GB at 4K context and 6.8GB at 100K context with 4-bit KV cache. These figures keep the footprint below the device’s physical limits, leaving room for the app and OS.

The implications go far beyond a successful experiment. First, it signals that the path of extreme quantization has not yet bottomed out: if an entire LLM, including embeddings and output head, can be reduced to 1 bit without catastrophic loss, then the gap between “server-grade” and “pocket-sized” models is narrowing rapidly. For those evaluating on-premise or edge deployment, the message is that computational power once confined to datacenters is beginning to fit on consumer hardware, with all the benefits in latency, privacy, and independence from connectivity.

Second, the fact that math withstands the compression better than general knowledge raises a design question: perhaps future on-device models will specialize in analytical and tool-like tasks, while information retrieval will still demand higher precision, possibly with controlled cloud offloading. A segmentation of inference by workload type is taking shape.

Third, the no-escape-hatch approach challenges the very architecture of compressed LLMs. If it works on a 27B Qwen base, it could work on other families, pushing the community to rethink where 16- or 8-bit weights are truly necessary. This could accelerate the development of specialized silicon for ternary or binary computation, reshaping the hardware supply chain.

The fact remains that an iPhone today runs a model that until recently was unthinkable on a mobile device. It is not a lab demo but a product distributed through the Atomic Chat app. The team behind the app confirms they worked to optimize execution on Apple Silicon using the MLX framework. It is a concrete taste of a future where high-level conversational AI can live entirely in your pocket, with no cloud intermediary. For companies investing in data sovereignty and local infrastructure, it is a signal that the frontier is shifting downward: ever more accessible hardware, ever more compressed models, and control returning to the hands of the device owner.