Trinity Large

398B total / 13B active MoE (256 routed + 1 shared expert, top-4 routing, 1.56% activation ratio). 60 transformer + 6 dense layers, 3072 model dim, 48 query heads, 8 KV heads. Interleaved local (4096 window) and global attention with gated attention and depth-scaled sandwich norm. Sigmoid routing with SMEBU load balancing.

Trained on 17T tokens in 3 phases using 2,048 NVIDIA B300 GPUs in ~33 days (~$20M total for all Trinity models). Muon optimizer, zero loss spikes. 512K native context (tested to 1M). Data curated by DatologyAI (8T+ synthetic).

MMLU: 87.2 (instruct), MMLU-Pro: 75.3, GPQA-Diamond: 63.3. Trinity-Large-Thinking: tau2-Airline 88.0%, AIME25 96.3%, SWE-Bench 63.2%, PinchBench 91.9% (#2 globally). #1 most-used open model in the US on OpenRouter. Apache 2.0.