Arcee AI, a 30-person San Francisco lab, has released Trinity-Large-Thinking—a 399-billion parameter text-only reasoning model under the fully open Apache 2.0 license, representing what may be the most significant American contribution to open source AI in years.
The timing is deliberate. Throughout 2025, Chinese research labs like Alibaba’s Qwen and Z.ai set the pace for high-efficiency Mixture-of-Experts architectures. However, as 2026 begins, those labs have shifted toward proprietary enterprise platforms, leaving a void at the high end of the open-weight market. Meta’s Llama division retreated from the frontier landscape following mixed reception of Llama 4. For developers who relied on the Llama 3 era of dominance, the lack of a current 400B+ open model created an urgent need for alternatives.
Engineering Through Constraint
Arcee AI differentiate itself through what CTO Lucas Atkins calls “engineering through constraint.” The company first made waves in 2024 after securing a $24 million Series A, then took a massive risk in early 2026: committing $20 million—nearly half their total funding—to a single 33-day training run for Trinity Large, utilizing a cluster of 2048 NVIDIA B300 Blackwell GPUs.
“The strength of the US has always been its startups, so maybe they’re the ones we should count on to lead in open-source AI,” said Clément Delangue, co-founder and CEO of Hugging Face. “Arcee shows that it’s possible!”
The Architecture: Extreme Sparsity
Trinity-Large-Thinking is noteworthy for the extreme sparsity of its attention mechanism. While housing 400 billion total parameters, its MoE architecture means only 1.56%—approximately 13 billion parameters—are active for any given token. This allows the model to possess the deep knowledge of a massive system while maintaining the inference speed and operational efficiency of a much smaller one, performing roughly 2 to 3 times faster than peers on the same hardware.
Training such a sparse model presented significant stability challenges. To prevent a few experts from becoming “winners” while others remained untrained “dead weight,” Arcee developed SMEBU (Soft-clamped Momentum Expert Bias Updates), ensuring experts are specialized and routed evenly across a general web corpus.
The Data Curriculum and Synthetic Reasoning
Arcee’s partnership with DatologyAI provided a curriculum of over 10 trillion curated tokens, expanded to 20 trillion tokens for the full-scale model—split evenly between curated web data and high-quality synthetic data. Unlike typical imitation-based synthetic data where a smaller model simply mimics a larger one, DatologyAI utilized techniques to synthetically rewrite raw web text to condense information, helping the model learn to reason over concepts rather than merely memorizing token strings.
Tremendous effort was invested in excluding copyrighted books and materials with unclear licensing, attracting enterprise customers wary of intellectual property risks associated with mainstream LLMs.
From Chatbots to Reasoning Agents
The defining feature of this release is the transition from a standard “instruct” model to a “reasoning” model. By implementing a “thinking” phase prior to generating a response, Arcee has addressed criticisms of its January “Preview” release, which sometimes struggled with multi-step instructions in complex environments.
The “Thinking” update enables what Arcee calls “long-horizon agents” that maintain coherence across multi-turn tool calls without degradation. This directly benefits Maestro Reasoning, a 32B-parameter derivative of Trinity already being used in audit-focused industries to provide transparent “thought-to-answer” traces.
Trinity-Large-Thinking is available now on Hugging Face under Apache 2.0, providing enterprises and developers with a rare American-made open weights model capable of competing at the frontier level while remaining fully customizable and deployable on private infrastructure.