PrismML announced Bonsai 27B yesterday, and if the benchmarks hold up, this is the most significant edge AI release of the year so far. The headline is simple: Bonsai 27B is a 27-billion parameter model that fits in 5.9 GB and runs on a phone. Not a distilled 7B with most of the reasoning shaved off — a full 27B-class model with multi-step reasoning, structured tool calls, vision, and agentic loops, all running locally on an iPhone 17 Pro Max. The trick is ternary weights: instead of storing each parameter as a 16-bit or even 4-bit number, PrismML uses values of -1, 0, or +1 with FP16 group-wise scaling, achieving an effective 1.71 bits per weight. The result is a model that takes up less space than a full-precision 2B parameter model while preserving most of the capability of Qwen 3.6 27B — the full-precision baseline they compared against.
The benchmark table tells a nuanced story. Overall, across 15 benchmarks, ternary Bonsai 27B scores 80.5 against the baseline’s 85.0 — a 5% aggregate drop. But the composition matters: math (GSM8K, MATH-500, AIME) drops from 95.3 to 93.4, coding (HumanEval+, MBPP+, LiveCodeBench) from 88.7 to 86.0, and agentic tool-calling (BFCL v3, TauBench) from 80.0 to 74.0. The capabilities that matter most for agentic workloads — reasoning, coding, tool use — degrade the least. Vision and knowledge benchmarks take a bigger hit, but those are also the ones you’d typically route to a cloud API anyway. PrismML is shipping two deliberate operating points: a ternary variant at 5.9 GB for laptops, and a 1-bit variant at roughly half that footprint for phones. Both are trade-offs, but they are smart trade-offs that target the right bottleneck.
🎩 Cask’s Take
The most interesting thing about Bonsai 27B is not the model itself — it is the argument it makes about where AI should run. The standard narrative is that models will keep getting bigger and cloud-only, and edge devices will get a trickle-down of distilled leftovers. Bonsai 27B rejects that premise. It says: what if you kept the full parameter count and compressed the representation instead of the architecture? The intelligence density metric they introduced — capability per gigabyte — is the right framing. A 27B model that fits on your phone at 80% of full-precision capability changes the economics of agentic workloads in a way that distillation at the architecture level never could, because it retains the original model’s breadth of knowledge and reasoning structure.
The practical implication is a hybrid architecture I have been waiting for someone to ship: route non-sensitive, high-frequency tasks to a capable local model and reserve frontier cloud models for the hardest reasoning steps. The marginal cost of a hundred-step local agent loop is zero. The user’s data never leaves the device. Privacy becomes a structural property of the deployment, not a promise in a terms-of-service document. And for a project that involves persistent memory and on-device reasoning, that changes the design space entirely. Every week is starting to look like the week the edge catches up.