400B total / 13B active (MoE)
Apache 2.0
Flagship Trinity family release with three checkpoints from the same run: Trinity-Large-Preview (lightly post-trained, chat-ready), Trinity-Large-Base (full 17T-token pretrained checkpoint), and Trinity-Large-TrueBase (10T-token pre-anneal checkpoint without instruct data). Release materials describe Trinity Large as a sparse MoE model with approximately 13B active parameters per token, using 256 experts with 4 active per token (4-of-256 routing, ~1.56% sparsity), plus a shared expert. Architecture and training notes reference AFMoE design, HSDP plus expert parallelism, and large-scale infrastructure using 2,048 NVIDIA B300 GPUs; official blog text states pretraining completed in 33 days and cites an all-in project cost around $20M. Data curation is attributed to DatologyAI with a 17T-token recipe across phased mixes (including code, STEM, reasoning, and multilingual content), with release messaging noting 14 non-English languages. Context support is presented as 512K native, with preview API operation at 128K under 8-bit quantized serving. Benchmark tables in release materials highlight strong base-model performance including MMLU 87.2 and competitive Math-500 positioning, while Preview is positioned for interactive agentic and coding-oriented deployment across open inference stacks and hosted endpoints.