Trinity Mini

Trinity Mini sits in the middle of Arcee AI's Trinity family of open-weight models, designed to balance enterprise utility with accessibility for tinkerers. It is built as a 26-billion-parameter sparse mixture-of-experts architecture with 128 experts and 8 active per token, yielding roughly 3 billion active parameters per forward pass. This MoE design lets the model carry broad world knowledge while keeping inference compute close to a much smaller dense model, an arrangement that the Arcee team highlights as central to the family's efficiency goals. The model is tuned for reasoning tasks, and in Arcee's own testing it produces comparable total token counts to leading instruction-tuned competitors, suggesting it reaches conclusions without bloated chain-of-thought overhead. Training for Trinity Mini drew on 10 trillion tokens curated in partnership with Datology, extending the dataset used for Arcee's AFM-4.5B checkpoint with additional mathematics and code material. The run was executed on a cluster of 512 H200 GPUs provided by Prime Intellect using hybrid sharded data parallelism. The resulting model exposes a 131,072-token context window with an equal maximum output length and is available open-weight, allowing direct inspection, fine-tuning, and self-hosted deployment beyond any hosted API. Together these choices position Trinity Mini as a practical middle ground for teams that want reasoning-grade quality without paying the compute cost of a full 26B dense model, and without surrendering control of the weights.

For practical fit, Trinity Mini suits workloads that need sustained reasoning, structured tool use, or long-context document analysis at a modest active-parameter footprint. Its open-weight release and Clarifai-hosted API both lower the barrier for teams who want to experiment locally before committing to a production stack, and the extended context window makes it viable for codebase review, multi-document synthesis, and agentic pipelines that accumulate large prompts. Developers integrating it through Clarifai or OpenRouter-compatible endpoints can pair it with temperature control and structured outputs for deterministic behavior in pipelines. Compared with larger dense reasoning models, Trinity Mini's MoE design aims to deliver similar answer quality at lower per-token inference cost, making it a sensible option for organizations standardizing on Arcee's Trinity lineup as their default reasoning backbone.