Companies that use services like ChatGPT or Claude via API do not actually own these models because they send their data to third parties, pay for tokens, and have no access to the internal workings of the model. This situation is a serious concern for banks, hospitals, and government institutions.
For this reason, Arcee AI announced it has raised $150 million in funding to develop open models that enterprises can run directly in their own data centers.
The San Francisco-based company Arcee AI announced its Series B funding round. The leaders of this round were Vista Equity Partners, Cambium Capital, and Emergence Capital, with participation from Microsoft M12's venture division, AI10 Ventures, Hitachi, IAG, P7, and Wipro.
This funding valued Arcee at over $1 billion, making it a new unicorn; after accounting for the new capital, the post-investment valuation reached approximately $1.15 billion.
In 2025, McQuade decided to invest in the company and build advanced models from scratch. He noted in an interview that the startup had about $30 million and decided to allocate 65–70 percent of these funds to training new models.
The result of this investment was the Trinity model line. Over a period of about six months, Arcee transformed from a dense model with 4.5 billion parameters into Trinity Large—a mixture of experts (MoE) model with 400 billion parameters and 13 billion active parameters per token.
Trinity Large was released in early 2026 as a highly efficient, permissively licensed model, fully developed in the United States. It is considered the first frontier-scale open model created end-to-end in America since Meta stopped releasing Llama models.
The most interesting aspect in the industry was the development cost. Arcee reported that the entire model line for 2025, including Trinity Large, was created for approximately $20 million. This amount covers salaries, computing power, data, infrastructure, and operating expenses.
In an industry where many believe that training a new model requires billions, $20 million is an astonishingly low figure. This common view was already being challenged when DeepSeek announced the creation of a top-tier model for less than $6 million.
Tests show that the Trinity models outperform Meta's Llama 3 and match the level of Mistral and leading Chinese models. This allows Arcee to become one of the few American labs capable of competing convincingly in the open-weight segment.
The startup plans to use the raised funds to accelerate the development of the next generation of the Trinity model family. This involves creating larger models, improving reasoning capabilities, and developing more efficient architectures that can run on less powerful hardware.
Another important direction is expanding collaboration with the U.S. Department of Energy and its national laboratories. Open models that users can deploy on private infrastructure are very attractive for governmental and research tasks where data cannot leave secure environments.
The third direction is product development. Arcee intends to create a new generation of tools to help organizations fine-tune, evaluate, implement, and operate open models. This includes tools for fine-tuning, security assessment, and deployment within the Vista Equity Partners portfolio companies. Vista owns over 90 corporate software companies, all of which are potential customers for the open models they can own and manage.
