We’re releasing Fantail, small language models tuned for coding agents with schema‑first prompting and constrained decoding. On Terminal‑Bench: 0.5B 28.4%, 1.3B 34.7%, 3B 41.3% vs GPT‑5 mini 39.8% and Sonnet 4.5 45.7% (3 seeds, no web). Reproducible runs + CSV→figure script included.

PS: we will release the weights and details later this week.

Links

Paper/overview: We will be posting here the details later this week. Weights: https://huggingface.co/autohandai/models Code: https://github.com/autohandai/fantail