Arcee AI Releases Trinity Large Thinking: Open-Source Reasoning Model

Arcee AI has officially released Trinity Large Thinking, an open-weight reasoning model distributed under the Apache 2.0 license. Designed to shift the open-source landscape toward complex, multi-step reasoning, the model is specifically engineered for long-horizon autonomous agents, multi-turn tool calling, and maintaining context coherence over extended workflows. By providing a transparent alternative to proprietary reasoning models, Arcee AI aims to support developers in building reliable, agentic systems.

Architecture and Efficiency

Trinity Large Thinking is a sparse Mixture-of-Experts (MoE) model featuring 400 billion total parameters. To ensure inference efficiency, the model employs a 4-of-256 expert routing strategy, activating only 13 billion parameters per token. This architectural choice allows the model to provide the world-knowledge density of a massive model while avoiding the latency typically associated with dense 400B architectures.
The model’s development involved several technical innovations, including the use of the Muon optimizer during its 17-trillion-token pre-training phase, which enhances capital and sample efficiency. Additionally, Arcee implemented SMEBU (Soft-clamped Momentum Expert Bias Updates), a load-balancing strategy designed to prevent expert collapse and ensure uniform utilization of the model's pathways. The architecture also incorporates interleaved local and global attention alongside gated attention to improve detail recall within large contexts.

Reasoning and Agentic Performance

A defining characteristic of Trinity Large Thinking is its internal "thinking" process, which occurs prior to delivering a final response. This mechanism allows the model to plan multi-step tasks and verify its logic before generating an output. This capability is critical for its primary use case: operating within complex software environments where reliability and instruction-following accuracy are paramount.
The model’s effectiveness is reflected in its performance on PinchBench, a benchmark focused on autonomous agent capabilities. Trinity Large Thinking currently holds the number two spot on the leaderboard, trailing only Claude Opus-4.6. To support these agentic loops, the model features a 262,144-token context window, enabling it to process massive datasets, complex codebases, and extended conversational histories without losing track of initial instructions.

Open Ownership and Deployment

By releasing the model under the Apache 2.0 license, Arcee AI provides developers and enterprises with the ability to audit, fine-tune, and self-host the model. This "true open" approach ensures that organizations can maintain data sovereignty and meet regulatory compliance requirements by keeping the model within their own infrastructure. The model is available on Hugging Face, and its extended context window is accessible via OpenRouter, facilitating its integration into diverse technical workflows.