Trinity-Large-Thinking: Open-source 398B MoE (13B active) for agentic tasks
a day ago
- #Open Source AI
- #AI Agents
- #Mixture-of-Experts
- Trinity-Large-Thinking is a Mixture-of-Experts (MoE) model with 398B total parameters but only 13B active during inference, enabling fast performance while leveraging extensive knowledge.
- It was pretrained on 17 trillion tokens and specifically post-trained on agentic tasks like tool-calling trajectories and multi-step reasoning, integrating reasoning and tool use from the start.
- The model maintains thinking tokens across the entire agent loop, preserving reasoning traces in context to inform decisions and avoid resetting memory between steps—key for multi-step tasks.
- Benchmarks show Trinity excels in agentic scenarios, outperforming models like Opus 4.6 on Tau2-Airline (88.0 vs. 82.0) and Tau2-Telecom (94.7 vs. 92.1), and scoring 98.2 on LiveCodeBench for coding tasks.
- Trinity requires significant infrastructure and is not for consumer GPUs; it's accessible via OpenRouter API or vLLM for custom deployments, with a strict note to preserve thinking blocks in context for effective operation.
- It is best suited for production agent systems (e.g., OpenClaw or Hermes Agent integration), not for general-purpose tasks, as it is specifically designed for multi-step agentic workflows.