Sonny was distilled from a larger 350M parameter teacher model using:
: The term appears in various TikTok contexts, often linked to fashion showcases, such as styling Ann Demeulemeester collections.
TinyModel Sonny demonstrates that a sub-15-million-parameter decoder-only model can deliver practical, real-time natural language understanding and lightweight generation on commodity microcontrollers. While not competitive with cloud LLMs on complex reasoning, Sonny excels in .
| Task | Metric | Sonny Score | Baseline (TinyBERT-4L) | |------|--------|-------------|-------------------------| | SST-2 (sentiment) | Accuracy | 81.2% | 83.4% | | CoLA (linguistic acceptability) | Matthews corr. | 0.52 | 0.58 | | RTE (entailment) | Accuracy | 64.3% | 66.1% |