PrismML Launches Bonsai 27B: Qwen3.6-Based, Claims On-Device iPhone 17 Pro, 90% Intelligence Retained
Startup PrismML launched Bonsai 27B, a multimodal model based on Qwen3.6, claiming it retains 90% of intelligence while running locally on a 12GB iPhone 17 Pro. It ships in 3-bit and 1-bit variants: the 3-bit uses ternary weights at 5.9GB and runs on ordinary laptops.
Bottom line
Startup PrismML launched Bonsai 27B, a multimodal model based on Qwen3.6. Its core selling point is aggressively-quantized on-device operation: claiming 90% intelligence retained, runnable locally on a 12GB iPhone 17 Pro.
Two quantized variants
Bonsai 27B ships in two versions, both trading ultra-low bits for size and memory:
- 3-bit variant: uses ternary weights, 5.9 GB, aimed at performance-minded users, runnable on ordinary laptops.
- 1-bit variant: more aggressive quantization, further compressing size and memory for runnability on memory-constrained devices like phones.
'Retains 90% intelligence' refers to capability retention on benchmarks versus the full-precision original—trading a 10% capability loss for the possibility of 'fitting into a phone'.
Why it matters: the core tension of on-device LLMs
Bonsai 27B touches the most fundamental tension in on-device LLMs—the tug-of-war between 'model capability' and 'device memory':
- A 27B model at full precision needs tens of GB of memory—phones simply can't hold it.
- Aggressive quantization (3-bit, 1-bit, ternary weights) is key to squeezing large models into phones, but lower bits mean greater precision loss.
- If Bonsai's claim (90% intelligence + phone-runnable) holds, it means the 'strong-yet-small' sweet spot is being pushed closer to the phone.
Ternary weights deserve a note
Ternary weights (each weight only -1/0/+1) are a hot direction in low-bit quantization—compressing storage and compute to the extreme while recovering precision via training/fine-tuning. Bonsai applies it in the 5.9GB 3-bit version, an engineering instance of this approach.
Caveats
'Retains 90% intelligence' and 'runs locally on iPhone' are PrismML's own claims. The real experience of aggressively-quantized models depends on task-specific precision loss, inference speed, heat, and battery—all needing third-party testing. Between an on-device model 'running' and 'being usable' there's often a sizable gap.