A 27B multimodal model squeezed to 1 bit, running on a phone

PrismML has released Bonsai 27B, a 27-billion-parameter multimodal model compressed hard enough to run on a phone. The trick is aggressive quantization: the 1-bit build stores each weight as either minus one or plus one, with a shared scaling factor per group, which brings the model down to about 1.1 bits per weight and 3.9 GB on disk. That matters because an iPhone 17 Pro gives a single app only around 6 GB of memory to work with, and the model has to share that space with its own KV cache and activations. A conventional 27B build never fit.

What makes the result interesting is how little breaks. Across 15 benchmarks the 1-bit version keeps about 90 percent of the full-precision model's score, and a slightly larger ternary build (three values per weight, 5.9 GB) keeps 95 percent. Math holds up best, dropping from 95.3 to 91.7. Tool-calling suffers most, falling from 80 to 66. The low-bit format runs through the entire network, including embeddings, attention, and the output head, with no higher-precision fallback layers. On an RTX 5090 the 1-bit model generates 163 tokens per second.

Why it matters

If you build mobile or offline apps, a capable multimodal model that fits in 4 GB means you can run vision and tool-calling on the device without a server. Test the tool-calling drop first, though, since that is where the 1-bit build loses the most accuracy and where agent workflows would feel it.