Arcee Trinity-Large-Thinking — 96% Cheaper Than Opus and Open Source
Arcee just released Trinity-Large-Thinking under Apache 2.0 and the numbers are hard to ignore. A 398B-parameter sparse MoE model with only 13B active parameters per token. It generates explicit reasoning traces in think blocks before responding — the same chain-of-thought architecture that made Opus 4.6 and o3 powerful. The price difference: $0.90 per million output tokens vs Opus 4.6's $25. That's 96% cheaper.
The agent angle is what makes this matter for the ecosystem. Trinity-Large-Thinking was specifically post-trained with agentic RL for long-horizon agents and multi-turn tool calling. This isn't a general reasoning model that you can also use for agents — it's a reasoning model designed for agents. The think blocks give you interpretable traces of the agent's decision-making, which is critical for debugging autonomous workflows.
The adoption numbers are already significant. Trinity Large Preview crossed 3.37 trillion tokens served on OpenRouter in its first 2 months and became the number one most used open model in the US. The weights are on Hugging Face under Apache 2.0 — no restrictions, no usage policies, no phone-home requirements. DigitalOcean is already offering hosted inference.
This is the rare US-made open-weight frontier model. VentureBeat called it exactly that. The open model landscape has been dominated by Meta's Llama (US, custom license), Alibaba's Qwen (China), Mistral (France), and DeepSeek (China). An American company releasing a frontier-class agent model under Apache 2.0 changes the geopolitical dynamics of the open model ecosystem.
https://www.arcee.ai/blog/trinity-large-thinking
← Back to all articles
The agent angle is what makes this matter for the ecosystem. Trinity-Large-Thinking was specifically post-trained with agentic RL for long-horizon agents and multi-turn tool calling. This isn't a general reasoning model that you can also use for agents — it's a reasoning model designed for agents. The think blocks give you interpretable traces of the agent's decision-making, which is critical for debugging autonomous workflows.
The adoption numbers are already significant. Trinity Large Preview crossed 3.37 trillion tokens served on OpenRouter in its first 2 months and became the number one most used open model in the US. The weights are on Hugging Face under Apache 2.0 — no restrictions, no usage policies, no phone-home requirements. DigitalOcean is already offering hosted inference.
This is the rare US-made open-weight frontier model. VentureBeat called it exactly that. The open model landscape has been dominated by Meta's Llama (US, custom license), Alibaba's Qwen (China), Mistral (France), and DeepSeek (China). An American company releasing a frontier-class agent model under Apache 2.0 changes the geopolitical dynamics of the open model ecosystem.
https://www.arcee.ai/blog/trinity-large-thinking
Comments