Physix Frontier · News Briefing Card (IT Home · Nov 6, 2026)

Taobao MNN open-sources 0.3B phone model, 350 token/s offline on iPhone 18 Pro

KEY FACTS

The Taobao MNN team has released and open-sourced MNN-Pocket-0.3B, a model with roughly 300 million parameters that focuses on Chinese and English support.
The model runs fully offline on the iPhone 18 Pro, with an average decoding speed of about 350 token/s.
The INT4 deployment file is 173MB, about 49% smaller than Qwen3-0.6B, with about 47% less peak memory overhead.
The model uses a 28-layer architecture, and its vocabulary has been trimmed from about 151K to 64K, a reduction of about 57%.
In tests on the OnePlus 13, a single-turn inference added about 2.7J of energy, and 100 tasks consumed about 0.3% of the battery.

KEY DATA

350token/siPhone 18 Pro decoding speed

173MBINT4 deployment size

约49%Deployment size reduction vs Qwen3-0.6B

73.20English IFEval score

PHYSIX OBSERVATION

The competitive focus for on-device models is shifting from parameter scale to usability per unit of resource. Taobao has compressed the model to 0.3B and tied it to its own inference engine, which both serves high-frequency calls inside the app and gives developers a low-barrier foundation. But whether 0.3B can support complex business scenarios remains to be verified; the real highlight is the effect of subsequent application integration.

Source: IT Home report