Arcee AI has released Trinity-Large-Thinking, a frontier open-weight reasoning model designed for complex, long-horizon AI agents. Available under the Apache 2.0 license, the model ranks second on PinchBench — just behind Anthropic's Claude Opus 4.6 — while priced at $0.90 per million output tokens, roughly 96% cheaper than its closest rival.
Key Highlights
- 400B sparse Mixture of Experts architecture with 13B active parameters per token, using 256 experts with 4 active per forward pass
- Second place on PinchBench, the benchmark measuring agentic task performance, trailing only Claude Opus 4.6
- Apache 2.0 license with open weights on Hugging Face, giving enterprises full ownership
- 96% cheaper than comparable frontier models at $0.90 per million output tokens
What Makes Trinity-Large-Thinking Different
Trinity-Large-Thinking builds on the foundation of Trinity-Large-Preview, which served 3.37 trillion tokens in its first two months on OpenRouter and became the most-used open model in the United States. The new release adds a reasoning layer — the model "thinks" before responding — which significantly improves its agentic capabilities.
Compared to the Preview version, Trinity-Large-Thinking delivers major improvements in multi-turn tool use, context coherence, instruction following, and stability across long-running agent loops. Arcee describes it as "the strongest open model ever released outside of China."
Built for Enterprise Agents
The model targets a specific gap in the market: enterprises that need frontier-level agent performance but want to own, inspect, and customize their models. The Apache 2.0 license allows organizations to post-train, host, distill, and deploy without restrictions.
"Developers and enterprises need models they can inspect, post-train, host, distill, and own," Arcee stated in the release announcement.
Availability
Trinity-Large-Thinking is available now through multiple channels:
- Arcee API at chat.arcee.ai
- Hugging Face with full open weights
- OpenRouter, where it is free for the first five days
The Preview model will remain available on OpenRouter with reduced hardware allocation.
Impact
The release intensifies competition in the open-weight model space, where Chinese labs like DeepSeek and Zhipu have dominated recent benchmarks. Arcee, a U.S.-based startup, positions Trinity as a domestically built alternative for organizations with data sovereignty or regulatory requirements.
For developers building agentic systems, the combination of frontier-level performance, permissive licensing, and aggressive pricing makes Trinity-Large-Thinking a compelling option — particularly for workloads that require long-running, multi-turn interactions with tool use.
Source: Arcee AI