Trinity (model family)

Trinity is the open-weight large language model family of Arcee AI, a United States company that describes itself as an open intelligence lab. Launched on December 1, 2025 with the small Trinity Mini and Trinity Nano Preview and extended on January 27, 2026 by the 400-billion-parameter Trinity Large, the family applies a single design philosophy at three scales: highly sparse mixture-of-experts routing, long context windows, and post-training aimed first at tool calling and agentic reliability rather than open-ended chat.¹² Press coverage framed Trinity Large as the largest open-weight model released by an American lab to that date, positioning the family as a US entry in an open-weight field otherwise dominated by Chinese releases.⁴⁵

Family tree

• AFM-4.5B (2025, first from-scratch Arcee foundation model)
  • Trinity Nano (December 2025, 6B total / 1B active, edge)
  • Trinity Mini (December 2025, 26B total / 3B active, agents)
  • Trinity Large (January 2026, 400B total / 13B active, flagship)
    • Trinity Large Thinking (April 2026, reasoning branch)

Design philosophy

Every Trinity model is a sparse mixture-of-experts Transformer with an unusually small active fraction. Trinity Nano activates roughly 1 billion of 6 billion total parameters, Trinity Mini about 3 billion of 26 billion, and Trinity Large about 13 billion of 400 billion per token, routing to 4 of 256 experts per layer. The intent is to buy the knowledge capacity of a large total parameter count while keeping per-token inference cost close to that of a small dense model, an economic bet on sparsity that the family shares with contemporaries such as DeepSeek and Kimi K3 but pushes further: Large's roughly 3 percent active ratio is among the sparsest of the 2026 flagships.²⁶ Nano and Mini carry 128K-token contexts; Large is trained to a 512K native context.

The second shared commitment is tool-use-first post-training. Across scales, Arcee describes curated pretraining data supplemented by heavy synthetic augmentation targeting tool calling, schema adherence, and error recovery, so that the models are optimized for function-calling agents and orchestration pipelines rather than benchmark chat.⁶ Trinity Large was trained from scratch on 17 trillion tokens in three phases (10T, 4T, and 3T), with data curation done alongside DatologyAI and more than 8 trillion synthetic tokens in the mix; the run used 2,048 NVIDIA B300 GPUs for 33 days at an all-in cost the company reported as around 20 million dollars, a notably lean budget for a 400B-class training run and a data point in ongoing arguments about scaling economics.²⁴

The family's from-scratch pedigree is recent. Arcee began as an enterprise small-model and model-merging company, absorbed the MergeKit toolkit, and built its SuperNova line by distilling Llama 3.1 405B before shipping AFM-4.5B, its first foundation model trained from scratch, in 2025. Trinity scales that recipe up by two orders of magnitude.¹⁴

Scales and release flavors

The three scales form a deliberate ladder: Nano for edge and embedded deployment, Mini for agents and tool orchestration at commodity cost, Large as the frontier-adjacent flagship. Alongside the scale ladder, the Trinity Large release introduced a three-flavor checkpoint ladder that functions as an open-science practice. ''Preview'' is a lightly post-trained instruct model in the ordinary sense of instruction tuning; ''Base'' is the best checkpoint after the full 17T-token recipe; and ''TrueBase'' is the early 10T-token checkpoint, released before any instruct data or learning-rate anneals touched the weights.² Publishing an unannealed, uncontaminated base checkpoint gives researchers a rare substrate for continued pretraining and post-training studies, and distinguishes the family from open-weight releases that ship only a final instruct model.

Reasoning branch and positioning

Trinity Large Thinking, released April 1, 2026 under the Apache 2.0 license, is the family's reasoning branch, following the test-time compute pattern established by o1 and DeepSeek-R1. Built on Trinity-Large-Base with supervised fine-tuning and reinforcement learning post-training on 1,152 H100 GPUs, it posted a developer-reported second place on PinchBench behind Claude Opus 4.6 at 0.90 dollars per million output tokens, which Arcee characterized as roughly 96 percent cheaper.³⁷

In the 2026 open-weight field, Trinity's closest framing is against Llama 4: Arcee reported Trinity Large Preview at 87.2 on MMLU and 24.0 on AIME 2025, roughly in line with Llama-4-Maverick Instruct, and TechCrunch cast the launch as a tiny startup building from scratch to best Meta AI's Llama.²⁴ Against the Chinese open-weight majors, the trillion-parameter-class Kimi K3 and GLM-5 remain larger and, on developer-reported numbers, stronger at the top end; Trinity's counterpositioning is provenance (US-trained on US infrastructure), permissive licensing, extreme sparsity, and the TrueBase disclosure ladder. VentureBeat framed the Thinking release as a rare powerful American-made open model that enterprises can download and customize.⁵⁷ Weights ship on Hugging Face, with serving through the Arcee API and OpenRouter.

See also

• Arcee AI
• Llama (model family)
• Kimi (model family)
• GLM (model family)

References