itstargetnews
Ai lab based in San Francisco Arcee done last year’s waves is one of the only US companies that trains major language models (LLM) from scratch and they are released under open or partially open source licenses to the public—enabling developers, solopreneurs, and even medium-to-large businesses to use powerful AI models for free and customize them at will.
Now Arcee is back again this week with the release of her biggest, most performant open language model to date: Trinity Greata 400-billion parameter mixture-of-experts (MoE), now in preview,
Along with the flagship release, Arcee sent a "raw" checkpoint model, Trinity-Large-TrueBasewhich allows researchers to study what a 400B small MoE can learn from raw data alone, before applying tuning and strengthening instruction.
By providing a clean slate at the 10-trillion-token mark, Arcee allows AI builders in highly regulated industries to perform real audits and conduct their own special alignments without inheriting the "black box" biases or formatting quirks in a general-purpose chat model. This transparency allows for a deeper understanding of the difference between a model’s intrinsic reasoning capabilities and the helpful behaviors dialed in during the final stages of post-training.
This launch comes as powerful Chinese open-source LLM alternatives from the likes of Alibaba (Qwen), z.AI (Zhipu), DeepSeek, Moonshot, and Baidu flood the market, effectively leading the category with high-efficiency architectures.
Trinity Large also comes after Meta has largely retreated from the frontier open-source scene. Following the April 2025 debut of Llama 4which was met by mixed receptionand former Meta AI researcher Yann LeCun later admitted that the company used several special versions of the model to increase scores on third-party benchmarks.
Amid this domestic vacuum, only OpenAI—along with it gpt-oss family released in summer 2025—and Arcee currently carry the mantle of new US-made open-source models completely trained from scratch.
As small as they come
The Trinity Large is notable for the extremely small size of its attention mechanism. An MoE architecture, "sting" refers to the model’s ability to selectively activate a subset of the total parameters for any given task.
While Trinity Large has 400B total parameters, only 1.56% (13B parameters) are active at any given time.
This architectural choice is important because it allows the model to have the "knowledge" in a large system while maintaining the inference speed and operational efficiency of a smaller one-achieving performance that is almost 2-3x faster than its peers on the same hardware.
Sovereignty and the "TrueBase" philosophy
The most important contribution of this release to the research community is Trinity-Large-TrueBase—a raw, 10-trillion-token checkpoint.
Unlike almost everyone "TOMORROW" release, which comes after "denied" by teaching tuning and reinforcing learning, TrueBase offers a unique, unadulterated view of foundational intelligence.
In the rush to create helpful models, most labs apply supervised fine-tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) before releasing the weights. Although this makes the model a better conversation, it can hide underlying knowledge distributions.
TrueBase provides a "AND base model" which have not gone through the learning rate anneals or the phase two and three pre-training where the instruction data is usually introduced.
For researchers and businesses in highly regulated industries, starting with TrueBase allows for true audits and custom alignments. As Lucas Atkins, Arcee’s CTO, said in a video call with VentureBeat: "Interestingly the checkpoint itself is already one of the best base models in the world".
Technology: engineering by constraint
The creation of Trinity Large is not the product of unlimited resources, but what Atkins calls it "engineering by prevention".
Trained for approximately $20 million in just 33 days, the model represents a masterclass in capital efficiency.
Arcee, a team of only 30 people, operates with a total capital of just under $50 million, making $20 million in training "back to the company" bet.
"I’ve always believed that having control, whether it’s financial or personnel or whatever, is very important to creativity," Atkins explained. "When you have an unlimited budget, you don’t have to engineer your way out of complex problems.".
Architecture: 4-of-256 Sparsity and SMEBU
Trinity Large uses a 4-of-256 bit MoE architecture, meaning it activates only 4 of its 256 experts for each token.
This high level of sparsity—one of the highest successfully trained—creates significant stability challenges during pre-training.
To solve this, Arcee created the Soft-clamp Momentum Expert Bias Updates (SMEBU). This mechanism ensures that experts are specialized and equally targeted in a general web corpus, preventing some experts from becoming "winners" while others remain untrained "dead weight".
The speed of the training flow was facilitated by Arcee’s early access to Nvidia B300 GPUs (Blackwell). These chips provide almost twice the speed of the previous generation of Hopper and a significant increase in memory.
"Pre-training is 33 days," Atkins announced. "We could have done it on Hopper, and it would probably take two to three months. And at that point, we’re into a whole new generation of models".
In collaboration with DatologyAIArcee uses over 8 trillion tokens of synthetic data. However, this is unusual "copying" synthetic data where a small model learns to speak like a larger one.
Instead, the intent is to take raw text on the web—such as blogs or Wikipedia articles—and synthetically rewrite it to condense the information into a smaller number of total tokens. This process helps the model learn to reason with the information instead of just memorizing the exact lines of tokens.
The architectural design also includes alternating local and global sliding window attention layers in a 3:1 ratio. This hybrid approach allows the model to be more efficient in long-term context scenarios. While trained for a 256k sequence length, Trinity Large natively supports 512k contexts, and evaluations suggest that it remains performant even at a 1-million-token horizon.
Technical comparison: Trinity Large vs. gpt-oss-120b
As an American alternative, Trinity Large can be compared to OpenAI’s gpt-oss-120b.
While both models use minimal architectures to achieve boundary-level performance under permitted licenses, they serve different operational roles.
While gpt-oss-120b currently contains only a handful of specific reasoning and math benchmarks, Trinity Large offers a significant advantage in context capacity and raw parameter depth for complex, multi-step agent workflows.
Sovereignty: filling the void
The release of Trinity Large is as much a geopolitical statement as a technical one. CEO Mark McQuade noted to VentureBeat in the same interview that the vacuum of American open-source models at the border level forced a pivot in Arcee’s strategy.
"This has become the kind of transition where players based in the US or the West have stopped opening these models," McQuade said. "We rely on these models to go to organizations and take them further … but the labs in China have started …".
For McQuade, this creates a confidence that American businesses are increasingly uncomfortable with. "Especially when we talk to large organizations, they are not able to use Chinese based architectures," he explained. "We want to be US champions. (It) actually doesn’t exist now".
By releasing under the Apache 2.0 license, Arcee provides the gold-standard permissive framework that allows companies to "YOURSELF" the model layer completely. This is important for industries such as finance and defense, where using a model hosted by a third party or a strict cloud provider is a non-starter.
Balancing intelligence with utility
Arcee now focused on the "current thinking model" to move Trinity Large from a general instruct model to a full reasoning model. The team wrestled with the balance between "intelligence versus utility"—trying to build a model that outperforms benchmarks is impossible "yappy" or inefficient in actual production applications.
"We built Trinity so you can own it," the team says, marking a return to the fundamental values of the American open-source movement. As the industry moves toward agentic workflows and more contextual requirements, Trinity Large positions itself not as a "wrappers," but as a sovereign infrastructure layer that developers can ultimately control.
