~398B total / ~13B active per token (Sparse MoE, AFMoE)
512K
Apache-2.0
Open-weight reasoning-optimized Trinity Large variant built from Trinity-Large-Base and post-trained with extended chain-of-thought and agentic RL for tool-use-heavy workflows. Official launch materials position Trinity-Large-Thinking as a frontier open reasoning model for complex long-horizon agents and multi-turn tool calling, and frame it as the official release that closes gaps left by Trinity-Large-Preview. Release claims emphasize stronger multi-turn tool use, better context coherence, cleaner instruction following, and more stable behavior in long-running agent loops, with native reasoning traces in <think></think> blocks that should be preserved in context across turns. Benchmark and deployment messaging highlights reported agentic performance including #2 on PinchBench and competitive scores across IFBench, GPQA-Diamond, AIME25, MMLU-Pro, and SWE-bench Verified-style evaluations, alongside API economics claims of about $0.90 per million output tokens. Architecture and training notes align with the Trinity Large stack: 256 experts with 4 active per token, 512K extended context, and large-scale training/post-training infrastructure, with deployment guidance for vLLM reasoning/tool-call parsers plus compatibility with OpenClaw and Hermes Agent frameworks.