Open and specialized foundation models broaden frontier claims

Mistral’s open-weight Large 4 claims 1T total parameters and 52B active parameters, while InstaDeep’s NTv3 describes large-scale genomic pretraining and experimentally validated enhancer design. Axion 1.0 adds an open agent-model recipe claiming to avoid several common training stages, but the available description lacks quantitative benchmarks. Reproducibility and independent matched evaluation remain necessary.