Arcee AI Challenges Claude Opus with Its Trinity-Large-Thinking Model

Arcee AI Challenges Chinese Giants with Trinity-Large-Thinking

Arcee AI recently unveiled Trinity-Large-Thinking, an open reasoning model aimed at competing with Claude Opus in agent tasks. To bring this project to fruition, the company allocated approximately half of its total venture capital.

In a sector dominated by Chinese players such as Qwen, MiniMax, and Zhipu AI, Arcee AI, an American startup, is trying to stand out with Trinity-Large-Thinking. This model, licensed under Apache 2.0, consists of around 400 billion parameters and is specifically designed for agent tasks. Thanks to a mixture of experts architecture, only 13 billion parameters are activated per token, allowing for efficient inference despite the model's massive size.

According to information provided by the company, the team trained the base model using 2,048 Nvidia B300 GPUs over a period of 33 days. This process cost approximately $20 million, representing nearly half of the funds raised by Arcee AI to date. Lucas Atkins, the CTO, described this model as the most powerful ever released outside of China in a blog post accompanying the launch.

Model Performance

Agent Benchmarks and General Reasoning

Trinity-Large-Thinking stands out for its ability to generate an explicit thought process, structured in blocks of reflection before each response. The model is optimized for tasks such as tool calling, multi-step planning, and autonomous workflows.

In agent benchmarks, Trinity-Large-Thinking performs at the level of Opus 4.6, particularly on Tau2 and PinchBench. However, it lags behind in general reasoning tests such as GPQA-D and MMLU-Pro.

Results on Hugging Face show that Trinity-Large-Thinking excels in agent benchmarks: it scores 88 on Tau2-Airline, ranks second on PinchBench with 91.9, just behind Claude Opus 4.6, which achieves 93.3, and reaches 96.3 on AIME25.

In contrast, for general reasoning, the scores are less impressive: GPQA-Diamond shows 76.3 and MMLU-Pro 83.4, while Claude Opus 4.6 scores 89.2 and 89.1, respectively.

Mixture of Experts Architecture

The model employs a mixture of experts architecture with 256 specialized sub-networks, but only four are active per token. This means that about 13 billion out of 400 billion parameters are used at each computation step, saving processing power without reducing the model's overall capacity. According to the technical report, the base model achieves competitive benchmark results with GLM 4.5, even though that model activates many more parameters per token.

To handle long texts, Trinity-Large combines two types of attention layers: local layers covering each section of the text and global layers extending over the entire context. This configuration supports long context windows without a proportional increase in computational costs. In practice, the model achieves a usable context window of 512K tokens, although it was trained with only 256K. During the Needle-in-a-Haystack test, which checks if a model can locate specifically placed information in long texts, it scored 0.976 at 512K.

Training Methods and Data

Custom Rebalancing Method

The early training phases encountered issues when individual experts collapsed. The distribution of tokens across the sub-networks drifted, with some experts no longer being utilized, and the model stopped improving. According to the technical report, the main cause was the existing load balancing method among the experts. It corrected imbalances with the same fixed step size each time, regardless of whether an expert was slightly or massively overloaded. With 256 experts, this created a constant oscillation that never reached a stable state.

The team developed SMEBU (Soft-clamped Momentum Expert Bias Updates) to address this issue, a new method that adjusts corrections proportionally to the actual deviation and smooths them over time. Combined with five other stabilization measures introduced simultaneously due to time pressure, this resolved the problem. Subsequently, the entire training phase remained stable without a single sudden spike in training loss. These spikes are a common and dreaded issue with large models, potentially ruining an entire training phase in the worst cases.

Synthetic Training Data

A significant portion of the training data is synthetic: over 8 of the 17 trillion tokens were generated by other AI models rather than retrieved from the web. This includes 6.5 trillion tokens of rewritten web text, approximately 1 trillion tokens of multilingual data, and around 800 billion tokens of code. The partner DatologyAI managed the data curation. According to the technical report, this ranks among the largest documented synthetic data generations for pre-training.

Prime Intellect provided the GPU clusters. Given that the B300 systems were brand new at the time, GPU errors occurred and could only be fixed through firmware updates.

The team also developed a new method for processing training data called Random Sequential Document Buffer (RSDB). Normally, particularly long documents can dominate several consecutive training steps and skew the data distribution. RSDB randomly shuffles the documents, which the technical report indicates significantly reduces fluctuations between individual training steps.

Adoption and Outlook

Early Adoption Despite Limited Fine-Tuning

After pre-training, the model underwent a second phase of fine-tuning focused on specific skills such as tool usage and multi-step tasks. According to the technical report, this phase, however, lasted less time than expected due to limited compute time on the GPU cluster. Arcee AI describes the current version as preliminary and plans for more extensive fine-tuning in the next iteration.

A previously released preliminary version operated on OpenRouter, where it processed 3.37 trillion tokens during its first two months. It was ranked among the most used open models in the United States on the platform, according to Arcee AI. The Thinking version is also online on OpenRouter and works with agent frameworks such as OpenClaw and Hermes Agent.

Shortly before Arcee AI's release, Google launched Gemma 4, a new family of open models also under Apache 2.0 license and partly built on a mixture of experts architecture.