- Who it helps
- Developers comparing compact Chinese-English language models
- Availability
- November 9 technical report, revised November 10. The abstract describes open-weight deployment; exact artifact terms and installation were not verified.
Treat the device measurement as a specific result
The authors describe a 300-million-parameter dense model with a 32,000-token context. They report a 173 MB quantized footprint and up to 466 tokens per second on an iPad Pro M5 using Metal.
Capacity and practical quality require separate checks
The work uses vocabulary pruning, distillation and quantization-aware training. Reported throughput does not measure concurrent-request memory, sustained device performance or suitability for a reader’s task.
What you can do
- Read the primary description and its stated limits before planning an evaluation. Open reference ↗
Where the evidence stops
Technical abstract checked; weights, device performance and license were not independently verified.
Sources read. Public source text checked November 14, 2026 UTC. No third-party project, model inference or full workflow executed.
Sources & further reading
Paper · paper
MNN-Pocket technical report ↗
November 9 paper, revised November 10. Reported 173 MB INT4 footprint and M5 throughput are deployment-specific.
MNN Team · page checked 2026-11-15
Coverage note: Technical abstract checked; weights, device performance and license were not independently verified.
The record over time (2)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Source recheck
MNNPocket-300M combines a small language model with training and runtime choices aimed at mobile devices.
- · Added to Current
MNNPocket-300M combines a small language model with training and runtime choices aimed at mobile devices.