Bonsai 27B: A 27B-Class model that runs on a phone

The Bonsai team announced Bonsai 27B, a 27-billion-parameter multimodal model based on Qwen3.6 27B that runs on a phone. The team said the model is the first of its capability class to run on a phone and extends earlier work showing that models with 1-bit and ternary weights can produce commercially useful language models. Bonsai 27B supports multi-step reasoning, structured tool calls, vision tasks and computer-use agentic loops. The team said a 27B model typically occupies roughly 54GB in 16-bit precision and even a 4-bit build at 18GB is too large for a phone. Bonsai 27B comes in two variants. Ternary Bonsai 27B uses ternary weights with FP16 group-wise scaling for a true 1.71 effective bits per weight and a size of 5.9 GB. The 1-bit Bonsai 27B uses binary weights with the same scaling for 1.125 effective bits per weight and a size of 3.9 GB, fitting within the memory budget of an iPhone 17 Pro. Both variants are multimodal with a vision tower in 4-bit form, carry a 262K-token context and support speculative decoding. The team said Ternary Bonsai 27B retains 95% of the full-precision baseline across a 15-benchmark suite and 1-bit Bonsai 27B retains 90%. The model runs natively on Apple devices via MLX and on NVIDIA GPUs via CUDA. Weights are available under the Apache 2.0 License.