A Chinese research team introduced Pinal, a protein‑design system that accepts natural‑language prompts and produces functional sequences through a two‑stage pipeline. The first stage uses T2struct, a text‑to‑structure module that converts a prompt into a token array encoding spatial relationships between amino acids and assigns probabilities to each arrangement. These tokens are then fed to SAProt‑T, which performs a global re‑evaluation to avoid discarding potentially superior functional sequences. Training leveraged three large protein databases—totaling 800 million keyword‑protein pairs—paired with AlphaFold‑predicted structures, feeding both tags and LLM‑generated plain‑language descriptions into the model. Pinal was benchmarked on four diverse targets: a fluorescent protein, two enzymes, and a complex‑assembly protein. For the fluorescent case, 50 000 designs were generated, filtered for GFP‑like barrel topology and a key modifiable amino acid, leaving 29 candidates; only one actually fluoresced after expression in *E. coli*. A similar workflow yielded low absolute success rates, but the approach expands the design space beyond known sequences. Separately, Hugging Face released Carbon X, an open‑weight Qwen3‑14B‑based model fine‑tuned for DNA using reinforcement learning and supervised fine‑tuning, achieving bilingual DNA‑English translation while running on consumer hardware (e.g., a MacBook). It competes with commercial models like GPT‑6.1 Sol but offers lower cost and local execution, though it struggles with precise positional information and can hallucinate with vague prompts. Both tools aim to make sophisticated biomolecular design accessible to biologists without deep bioinformatics expertise.