The MNN team from Taobao has released the MNN-Pocket-0.3B model to the public. This language model, containing approximately 300 million parameters, is designed to function directly within mobile applications. According to a report by IT Home on November 6th, the model is built on the MNN inference engine and is intended for large-scale universal service and e-commerce applications, such as Taobao, where limited computational power, memory, and battery life on the device must be distributed across numerous functions.
The team set four key goals for their model, which they termed an application-native large model. These goals included achieving a decoding speed of over 200 tokens per second on specific devices. Peak memory consumption had to remain below 300 MB, leaving a buffer of about 500 MB within the overall model budget for context growth and integration overhead. Furthermore, a single typical inference process should not consume more than one-hundredth of the reference phone's battery charge, meaning 100 runs would cost less than 1%. Finally, when performing specified business tasks, the model must demonstrate results comparable to 8B class models.
Testing showed that MNN-Pocket-0.3B operates fully offline on the iPhone 18 Pro using the MNN engine, achieving an average decoding speed of about 350 tokens per second. The INT4 deployment package, which includes weights, tokenizer, computation graph, and inference configuration, is 173 MB in size. On the iPad Pro with the M5 chip and using the Metal backend, the team recorded a decoding speed of 466 tokens per second and a prefill speed of 8721 tokens per second, with peak memory increasing by 228 MB relative to the baseline. On the Snapdragon 8 Gen 5 phone using the CPU backend, 251 tokens per second was recorded, and the additional energy cost for 100 runs on the OnePlus 13 was about 0.3% of the battery charge.
The model was trained from scratch, with a particular focus on Chinese and English languages. After testing dense configurations ranging from 50 million to 1.3 billion parameters, the team determined that the size around 300 million was the largest that simultaneously met memory and speed requirements. It was decided to use 28 layers with grouped-query attention, as transitioning to 36 layers did not increase average performance but slowed down decoding. The vocabulary was reduced from approximately 151,000 tokens from the Qwen3 tokenizer to 64,000, which allowed for reduced embedding and output layer costs.
The team reports an IFEval score of 73.20, a Chinese IFEval of 74.59, and an average of 50.00 across 11 Chinese and English benchmarks, surpassing similarly sized and larger models, including Qwen3-0.6B. The model card notes that, like other models with fewer than a billion parameters, it is not recommended for complex mathematical tasks or long-horizon code generation.
Variants published on Hugging Face include a base model for continued pre-training, a supervised fine-tuning checkpoint, an instruction-tuned model, a GGUF build, an MNN INT4 build, and a decision-making head model for candidate selection. The release also contains a draft speculative decoding model, DFlash2, which, according to the team, can double decoding throughput in certain tasks.
