Bonsai 27B Runs on Smartphones: The First 27 Billion Parameter Model for Mobile
PrismML has announced Bonsai 27B, the new flagship model of the Bonsai family based on Qwen3.6 27B. It is the first model in its class of capabilities, 27 billion parameters, to run on a smartphone. The problem that the company claims to have solved is purely dimensional. A 27B model occupies about 54 GB at 16-bit precision, and even 4-bit compression remains above 18 GB: too much for any phone and for many laptops on the market.
Bonsai 27B comes in two variants. The first, Ternary Bonsai 27B, uses ternary weights (-1, 0, +1) with FP16 group scaling, achieving an effective density of 1.71 bits per weight and a footprint of 5.9 GB: this is the quality-oriented version, designed to run on a regular laptop while maintaining the full set of reasoning capabilities, tool-calling, and agentic behavior.
The second, 1-bit Bonsai 27B, pushes the compression to binary weights (-1, +1), also with the same group scaling, reaching 1.125 effective bits per weight and just 3.9 GB in total weight. It is the variant built to fit within the memory budget of an iPhone 17 Pro, and it is the first to bring a 27B class model to a phone.
Bonsai 27B runs locally on smartphones and has 27 billion parameters. The low-bit compression, the company explains, traverses the entire network: language network, embedding, attention, MLP, and LM head, with no shortcuts to higher precision at any point. Both variants remain multimodal, with the visual part compressed to 4 bits to continue reading screenshots, documents, and camera inputs. The context reaches 262,000 tokens, with speculative decoding support to accelerate generation without loss of quality. The model weights are available today under the Apache 2.0 license.
On a suite of 15 benchmarks covering knowledge, reasoning, math, programming, instruction following, tool calling, and vision (evaluated in extended reasoning mode), Ternary Bonsai 27B maintains 95% of the performance of the full precision model, while 1-bit Bonsai 27B stops at 90%.
Looking at individual categories instead of averages, math and programming remain almost intact, and tool calling stays just a few points shy of the original model, precisely the capabilities on which an agentic workflow relies. PrismML claims that the most aggressive low precision build available conventionally on the same base model scores significantly lower than 1-bit Bonsai 27B, while occupying 2.5 times more memory.
In terms of speed, the company states up to 163 tokens per second in 1-bit and 134 tokens per second in ternary on an NVIDIA GeForce RTX 5090, while on a Mac with the M5 Max chip it drops to 87 tokens per second in 1-bit and 58 in ternary.
However, the real constraint is not the file size but the memory actually available: an iPhone with 12 GB of RAM allocates about 6 GB to apps, a budget that the model must share with KV cache and activations. No conventional build of a 27B model comes close to that threshold; at about 4 GB, 1-bit Bonsai 27B is the first to fit within it with operational margin, which explains the choice to offer two distinct variants instead of a single compression.
PrismML (you've heard of it here before) claims these results as a continuation of a path already started with previous models of the Bonsai family, at 1 bit and ternary, and ties them to a broader argument: the local execution of agents capable of performing prolonged tasks, without every intermediate step, screenshot, or private document having to pass through a cloud provider's servers. The model is born from a team of researchers at Caltech, with support from Khosla Ventures, Cerberus, and Google, and the backing of Samsung.