淘宝 MNN 团队开源 0.3B 端侧模型，实现超高解码吞吐

On November 6, 2026, the MNN team at Taobao released MNN-Pocket-0.3B, an open-source on-device large model for smartphones, aimed at the task demands and resource budgets of mobile apps. The stated goal is to balance capability, response speed, and sustained running experience under limited compute, memory, and energy.

For comprehensive phone apps like Taobao's own, the team frames what it calls a 'native large model for phone apps' through four design targets. On speed, decoding throughput should exceed 200 token/s under the specified device and context configuration, benchmarked against the generation throughput of fast cloud model services, so that local generation feels smooth. On memory, peak memory should stay below 300MB under the stated test conditions; the team notes that even on devices with 8~16GB of memory, that capacity is shared by the system and multiple apps, so keeping inference peaks under 300MB against an approximate 500MB model-module budget leaves headroom for growing contexts and integration overhead. On power, under a typical inference load (512~1024 input tokens and 128 output tokens), the incremental energy per inference should not exceed one ten-thousandth of the reference phone's battery energy, meaning 100 inferences draw no more than 1% of total battery energy, to control drain from frequent calls. On capability, the aim is to match the results of an 8B-class model in specified business scenarios, covering Chinese and English understanding, world knowledge, math, and code reasoning, with emphasis on instruction following, and with long-context and multi-step tasks in mind.

The available reporting stops at these design targets; no further details were published on access to the code and weights, or on the specific test device models.