MNN-Pocket-0.3B
Alibaba's Taobao MNN team open-sourced MNN-Pocket-0.3B, a 300M-parameter on-device language model with a 173MB INT4 package that decodes about 350 tokens per second on an iPhone 18 Pro.
The model was trained from scratch for Chinese and English and sized at about 300M parameters, the largest the team found that met its memory and speed targets. It uses 28 layers with grouped-query attention and a vocabulary cut from about 151,000 to 64,000 tokens. The release includes base, SFT, instruct, GGUF, INT4 MNN and decision-head variants plus a DFlash2 speculative decoding draft model.
- Date
- Friday 6 November 2026
- Lab
- Taobao MNN team (Alibaba)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Decode speed, iPhone 18 Pro | about 350 tokens/s offline via MNN engine | company |
| INT4 package size | 173MB weights, tokenizer, compute graph and config | company |
| Decode speed, iPad Pro M5 (Metal) | 466 tokens/s prefill 8,721 tokens/s, peak memory +228MB | company |
| Decode speed, Snapdragon 8 Gen 5 CPU | 251 tokens/s | company |
| Battery cost, OnePlus 13 | about 0.3% for 100 runs | company |
| IFEval | 73.20 Chinese IFEval 74.59 | company |
| Average over 11 Chinese and English benchmarks | 50.00 team says ahead of same-size and larger models including Qwen3-0.6B | company |
Figures come from the team's own tests as relayed by Pandaily citing IT Home. The license was not stated.
Sources
This record was checked against its sources on 6 November 2026. How we check