AI-generated concept illustration: a paper-circuit garden inside a transparent phone-shaped container represents compact local AI.
Keeping a model resident on a phone usually runs into memory and wait time first. Alibaba's MNN team has made public MNN-Pocket-0.3B, which optimizes the model and inference engine together: about 300 million parameters, supports Chinese and English, and the quantized MNN build is 173MB. The official repository records this release as November 4, 2026; media reports appeared on November 6, and that cannot be used to push the release date back by two days. [S1][S2][S3]
As of November 8, the official model page provides download and run entry points, under the Apache 2.0 license. In addition to the INT4 version for MNN, the project lists resources including the base model, intermediate supervised fine-tuning versions, and GGUF; developers can also use the model through Transformers. The model card gives a deployment context limit of 32K. The news here is the available model and deployment assets, not a phone assistant already launched for all users. [S1][S2]
The official repository discloses that decoding throughput tested on iPhone 18 Pro is 350 tokens per second. This is a result the developer provided on a specific device and configuration; AIYXL has not reproduced it on the device. The 173MB describes the quantized build size and cannot be directly treated as the full runtime memory footprint of an application, nor can it be used to infer that all phones have the same speed. [S1][S2]
From an application engineering perspective, the value of this kind of release lies in letting teams establish their own resource ledger. Download size determines distribution burden, runtime memory affects the room available to coexist with other features, and response latency relates to whether an interaction is worth keeping local. Only by measuring these quantities separately is it possible to judge which part of the work the model is suited to handle.
A more practical testing approach is to first choose tasks with clear boundaries and easily verifiable results, such as converting short text into structured fields. In addition to recording average latency, testing should cover long inputs, consecutive calls, error recovery, and low-battery conditions. If the model can only give an answer quickly in a single demo, that is not enough to prove it is suitable for an application's daily load.
Open weights allow these questions to be further tested by external developers. What will truly determine integration value will be results under the target phone, target task, and sustained operation: how much waiting is saved, how many resources are consumed, and how the application catches errors after they occur.
Sources:
[S1] MNN official repository — the official log specifies November 4; parameters, license, and tests on the specified device
[S2] Official model card — weights, deployment versions, context, and framework support; developer's own account
[S3] IT之家, November 6 — media follow-up; its “released today” differs from the official log date, and the official log takes precedence
