Arcee AI is an American artificial intelligence company, self-described as an "open intelligence lab," founded in 2023 and headquartered in Miami, Florida.1 Its trajectory through the large language model industry was unusual: it began as an enterprise vendor of small, domain-adapted models assembled by merging and distilling other labs' open weights, then pivoted to pretraining its own foundation models from scratch. That pivot culminated in the Trinity family and the January 2026 release of Trinity Large, a 400-billion-parameter sparse mixture-of-experts model that press coverage described as the largest open-weight model released by a United States lab to that date.23
Stated mission and record
Arcee AI's homepage carries the tagline "Building Open Intelligence," glossed on the same page as: "AI works best when people can build on it freely. Arcee develops open-weight models in the U.S. and releases them for everyone to use."8 The company's about page adds that progress will be decided by "efficiency, adaptability, and the ability to own what you build," and pledges that "Every model we ship is permanent and portable, so there's no lock-in and no forced upgrades."9 As of July 2026 Arcee has published no separate formal charter or safety framework; its public positioning centers on making open weights, in its words, "the durable standard" for American AI.9
Much of the release record matches that positioning: the Trinity line, including Trinity Large and the Apache 2.0-licensed Trinity Large Thinking, is distributed through Hugging Face and OpenRouter alongside the company's own paid API, and AFM-4.5B's weights are likewise published on Hugging Face. Other parts of the record are narrower than the stated ideal: AFM-4.5B's weights initially shipped under Arcee's own model license before being opened under Apache 2.0, the SuperNova 70B distillation launched in September 2024 as a commercial product before being opened under Meta AI's more restrictive Llama license, MergeKit moved to the copyleft LGPLv3 in October 2025, and the openness Arcee describes applies to model weights rather than to training corpora, which for Trinity were curated in part with the external partner DatologyAI and had not been publicly released as of July 2026.
Origins: model merging and distillation
Arcee AI was founded in 2023 by Mark McQuade, Jacob Solawetz, and Brian Benedict, and initially sold enterprises small language models adapted to their private data. The company raised a $24 million Series A led by Emergence Capital in 2024 on that thesis, at a time when most attention (and capital) flowed to ever-larger closed models.1
Its early technical identity was inseparable from model merging: the practice of combining the weights of multiple fine-tuned Transformer checkpoints into a single model without additional training. MergeKit, the open-source toolkit created by Charles Goddard that became the community standard for merging checkpoints on Hugging Face, was absorbed into the company when Goddard joined; Arcee later relicensed MergeKit under the LGPLv3, effective October 31, 2025.4 Merging gave the company a way to produce competitive specialized models at a fraction of training cost, and merged variants of open bases such as LLaMA-derived checkpoints populated community leaderboards through 2024.
The complementary pillar was distillation. Arcee's SuperNova line, launched in September 2024, used offline logit distillation to compress Meta AI's Llama 3.1 405B into a 70B student model, initially offered commercially and later opened under the Llama 3 license; a 14B sibling, SuperNova-Medius, followed the same playbook at smaller scale.5 Together, merging and distillation positioned Arcee as a lab that built on top of the open ecosystem rather than contributing base models to it.
The pivot to pretraining: AFM-4.5B
That positioning changed in mid-2025 with AFM-4.5B, the first Arcee Foundation Model and the company's first model pretrained from scratch, announced in June 2025 with open weights following on Hugging Face that July. The roughly 4.5-billion-parameter model was trained on 8 trillion tokens, split between 6.5 trillion tokens of general pretraining and 1.5 trillion tokens of midtraining weighted toward mathematics and code, with a documented context extension to 64K tokens.5 AFM-4.5B was aimed squarely at enterprise deployment, marketed on performance, regulatory compliance, and cost rather than leaderboard supremacy, and a strategic funding round announced around the AFM platform financed the company's growing compute ambitions. In retrospect, AFM-4.5B was a rehearsal: proof that a small team could run a modern data-curation and pretraining pipeline end to end before betting the company on a far larger run.
Trinity: the frontier bid
The Trinity family, Arcee's open-weight sparse mixture-of-experts line, is that bet. It spans three sizes: Trinity Nano (6B total parameters, roughly 1B active, 128K context) for edge and embedded use, Trinity Mini (26B total, about 3B active) for agents and tool orchestration, and Trinity Large (400B total, roughly 13B active per token through 4-of-256 expert routing, 512K native context).6 Training combined curated web data with heavy synthetic augmentation targeting tool calling, schema adherence, and error recovery.
Trinity Large, announced January 27, 2026, was pretrained from scratch on 17 trillion tokens in three phases (10T, 4T, 3T) across 2,048 NVIDIA B300 GPUs in 33 days, with data curation performed in partnership with DatologyAI, including over 8 trillion synthetic tokens; the all-in cost was reported at around $20 million.2 Arcee shipped three flavors (a lightly instruction-tuned Preview, a Base checkpoint from the full recipe, and an early "TrueBase" checkpoint free of instruct data), with developer-reported scores of 87.2 on MMLU and 24.0 on AIME 2025, roughly in line with Llama 4 Maverick Instruct.2 TechCrunch headlined the release as a tiny startup building a 400B open model "to best Meta's Llama."3
A reasoning variant, Trinity Large Thinking, followed on April 1, 2026 under the Apache 2.0 license, built on Trinity-Large-Base with supervised fine-tuning and reinforcement-learning post-training on 1,152 H100 GPUs. Arcee reported it ranking second on PinchBench behind Claude Opus 4.6 while costing $0.90 per million output tokens, roughly 96 percent less; VentureBeat framed it as a rare powerful US-made open model that enterprises could download and customize.7 Both Trinity Large models are distributed through Hugging Face and OpenRouter alongside Arcee's own API.
Position in the field
Trinity's significance in the 2026 landscape was less about any single benchmark than about who released it and for how much. By early 2026, the open-weight frontier had been dominated by Chinese labs such as DeepSeek, Moonshot AI (Kimi K3), and Zhipu (GLM-5), following the trail blazed by DeepSeek-R1, while Meta AI's Llama 4 generation had disappointed many of its users; American open weights were increasingly scarce above the mid-size tier. A lab of a few dozen people pretraining a competitive 400B mixture-of-experts model for roughly $20 million demonstrated how far the cost floor had fallen: B300-generation compute efficiency, sparse-activation scaling, mature distributed training stacks, and outsourced data curation through partners like DatologyAI together put frontier-adjacent pretraining within reach of startup budgets.23 In the open-versus-closed debate, Arcee occupies a role sometimes compared to Mistral AI's in Europe: a small, capital-efficient national champion arguing that open weights and commercial viability can coexist.
See also
References
-
Refresh Miami, "Arcee AI secures $24M Series A to transform the landscape of small language models," 2024; company profile data as of July 2026. ↩↩
-
TechCrunch, "Tiny startup Arcee AI built a 400B-parameter open source LLM from scratch to best Meta's Llama," January 2026. ↩↩↩
-
Goddard, C., et al., "Arcee's MergeKit: A Toolkit for Merging Large Language Models," arXiv:2403.13257, March 2024; Arcee AI, MergeKit relicensing notice, October 2025. ↩
-
Arcee AI, "Meet Arcee-SuperNova," September 2024; Arcee AI, "Announcing Arcee Foundation Models," June 2025. ↩↩
-
Arcee AI, Trinity family overview (arcee.ai/trinity), 2026. ↩
-
Arcee AI, "Trinity Large Thinking," April 2026; VentureBeat coverage, April 2026. ↩
-
Arcee AI, "Building Open Intelligence" (homepage), official website, as of July 2026. ↩
