PrismML fits a 27-billion-parameter AI model into 3.9 GB — small enough for an iPhone

PrismML has released Bonsai 27B, a model based on Qwen3.6 27B that represents what the company calls the first 27-billion-parameter AI to fit comfortably on a phone. The achievement comes through extreme weight quantization: instead of storing each parameter as a 16-bit float, Bonsai 27B uses binary {−1, +1} weights — just one bit per parameter — to compress a model that would otherwise consume 54 GB of memory down to 3.9 GB, small enough to run within the memory budget of an iPhone 17 Pro.

Until now, the 27B parameter tier has been solidly in laptop-or-server territory. Even aggressive 4-bit quantization of a 27B model yields around 18 GB — too large for any current smartphone. PrismML's approach pushes the compression much further without catastrophic quality loss, building on techniques the company demonstrated with earlier 1-bit and ternary releases at smaller scales.

Two variants, two use cases

Bonsai 27B ships in two configurations:

• Ternary Bonsai 27B (5.9 GB): Uses {−1, 0, +1} ternary weights at 1.71 effective bits per weight. This is the quality-oriented variant and targets everyday laptops, retaining 95% of full-precision performance across PrismML's 15-benchmark suite.
• 1-bit Bonsai 27B (3.9 GB): Uses binary {−1, +1} weights at 1.125 effective bits per weight. This variant targets phones and fits within iPhone 17 Pro's memory budget, retaining 90% of full-precision performance.

Both variants include multimodal support, with the vision tower shipped in a compact 4-bit form. They carry a full 262K-token context window, support speculative decoding for faster inference, and handle multi-step tool calls and agentic loops — the kind of sustained workloads where on-device models have historically stumbled.

The benchmark picture

PrismML benchmarked both variants across 15 tests in thinking mode. Math and coding scores are nearly unaffected: the ternary model scores 93.4 on math versus the baseline's 95.3, and the 1-bit model holds 91.7. Tool-calling and agentic tasks show more compression impact — the 1-bit model scores 66.0 versus the baseline's 80.0 — but PrismML notes these still outperform conventional 4-bit builds of the same base model while occupying 2.5× less memory.

The company frames this in terms of "intelligence density" — performance per gigabyte — where 1-bit Bonsai 27B achieves 0.53 per GB, roughly 10× the full-precision baseline and 2.7× the best conventional low-bit alternative.

Why it matters for agentic AI

The practical upshot is that tasks requiring a capable reasoning model — document analysis, structured data extraction, multi-step code generation — can now run entirely on device, without sending data to a cloud API. That matters for privacy-sensitive workflows, latency-critical applications, and any scenario where network connectivity is unreliable.

Bonsai 27B is available now under the Apache 2.0 license, making it free for commercial use. As reported by PrismML, this is a milestone in the continuing push to make large-model capability available without cloud infrastructure.