Qwen3.6-27B Now Runs on an iPhone, at 1.125 Bits Per Weight
A twenty-seven-billion-parameter model just ran on an iPhone, at eleven tokens a second. It's Bonsai 27B, from Caltech spinout PrismML, and the weights are Apache-2.0.
Qwen3.6-27B, converted post-hoc into two builds
The base is Alibaba's Qwen three-point-six twenty-seven B, converted after training into two builds. Ternary runs on laptops. The one-bit build is three point nine gigabytes, the only one that fits an iPhone seventeen Pro Max's per-app memory budget. The size isn't the story; what it kept is.
A pretrained 27B, converted to true 1-bit, every layer
Nothing this big has run at true one bit before; the BitNet line needs from-scratch pretraining and stops around two billion parameters. Even the conventional two-bit quant really spends two point eight bits per weight. Bonsai converts every layer of the already-trained language model; the method stays proprietary.
Ternary keeps 94.6% of full precision, in 5.9 GB
On PrismML's own identical-harness eval, fifteen thinking benchmarks, the ternary build keeps ninety-four point six percent of full precision in five point nine gigabytes. The conventional two-bit quant keeps eighty-five and a half percent, in nine point four gigabytes. That's smaller, and better.
The "2-bit" quant collapses on reasoning; Bonsai holds
The average hides where cheap quants die. On AIME twenty-six, same harness, the two-bit quant collapses to fifty-seven point five. Ternary Bonsai holds eighty-seven point five. Even the one-bit build holds eighty-seven. That same quant still scores eighty-nine on MMLU-Redux. A casual test would miss the damage.
672 tokens per one percent of battery
On the phone it's not a stunt. One percent of battery buys six hundred and seventy-two tokens, about sixty-seven thousand on a full charge.
The fine print: PrismML's harness, PrismML's forks
Here's the fine print. Every quality number is PrismML's own run, baselines included, one harness; no third party has evaluated the twenty-seven B yet. The steepest cost is agentic tool use: tau-squared-Bench falls from eighty-three to sixty-one at one bit, and agentic coding isn't targeted. And today it runs only on PrismML's forks. LM Studio fails.
Apache-2.0, ungated, on Hugging Face now
huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf
Together AI: ternary hosted free · whitepaper: github.com/PrismML-Eng/Bonsai-demo
Both builds are on Hugging Face, Apache-2.0, ungated. Together AI hosts the ternary free. If reasoning at one point one two five bits per weight survives independent evals, the deployment math for local AI just changed.
