# Qwen 27B model runs on iPhone 18 Neural Engine at 15–70 tokens per second

A developer reports running a 27B-parameter Qwen model entirely on the iPhone 18's Neural Engine, reaching 15–70 tokens per second using aggressive 1.5-bit-per-weight quantization. The result, shared on Hacker News, suggests large models can increasingly run locally on phones rather than in the cloud. If it holds up, on-device inference at this scale could cut latency and cost while keeping user data private.

**Importance:** 3/5

## Sources

### Technology
- [Hacker News](https://twitter.com/anemll/status/2111517143087067309) — Sat, 17 Oct 2026 18:10:21 +0000