PrismML released Bonsai 27B, a compressed build of the Qwen3.6 27B model. The company says it is the first model in this size class that fits on a smartphone, shipping in two forms: a 5.9 gigabyte ternary version aimed at laptops, and a 3.9 gigabyte 1-bit version that PrismML says clears the memory ceiling on an iPhone 17 Pro. Both numbers are PrismML’s own, and the announcement contains no independent benchmark or third-party device test to confirm them.
The compression method is unusually aggressive. Ternary Bonsai 27B stores weights as {-1, 0, +1} values, a design PrismML calls 1.71 effective bits per weight. The 1-bit version drops to binary {-1, +1} weights, or 1.125 effective bits per weight. PrismML says both variants apply this low-bit scheme across the entire network, including attention layers and the output head, rather than reserving a few layers for higher precision, which is the more common approach among other compressed open models.
Scale matters here because of the arithmetic PrismML lays out: a 27-billion-parameter model normally needs about 54 gigabytes at 16-bit precision, and even a standard 4-bit version lands around 18 gigabytes, too big to fit a phone’s usable memory pool. PrismML estimates a 12 gigabyte iPhone exposes only around 6 gigabytes to any single app once the operating system and the model’s own cache take their share. At 3.9 gigabytes, Bonsai 27B is small enough to clear that ceiling with room left over, according to the company.
On accuracy, PrismML reports its ternary variant holds onto roughly 95% of the full-precision model’s score across a 15-benchmark suite covering math, coding, tool calling, and vision, while the 1-bit version keeps about 90%. Again, these figures come exclusively from PrismML’s own testing. No outside lab, leaderboard, or reviewer appears in the announcement to corroborate them.
The more consequential claim is architectural, not just about file size. Agentic systems, ones that call a model dozens or hundreds of times to complete a task, are expensive and slow when every call is a network round trip. Each step costs tokens, and each step sends the user’s files, screen contents, or documents across the internet. A model that runs entirely on a phone or a laptop removes both costs: the marginal price of the hundredth tool call is zero, and nothing about the task leaves the device.
That is a different pitch than the on-device features Apple and Google have shipped over the past two years, which mostly handle narrow jobs like autocomplete or summarization. PrismML is proposing a 27-billion-parameter model that reasons, calls tools, and processes images locally, categories of work usually reserved for cloud-hosted frontier models.
PrismML reports throughput up to 163 tokens per second on an Nvidia RTX 5090 and 87 tokens per second on Apple’s M5 Max, both far more powerful than a phone chip, without publishing equivalent phone-side speed figures. Weights are available now under an Apache 2.0 license.
Teams evaluating on-device agents should treat PrismML’s benchmark table as a starting hypothesis, not a verdict, and run their own multi-step tool-calling test on the actual hardware before committing a 2026 product roadmap to local inference.
PrismML announced Bonsai 27B in a company blog post published July 14, 2026.
