A new sparse heavyweight
The latest update to the leading open-weight models for enterprise use reveals an exciting architectural shift. Arcee AI has released Trinity Large, a sparse mixture-of-experts model with roughly 400 billion parameters in total, of which only 13 billion are active per token. According to Layer3 Labs, the model was trained on 2,048 Nvidia Blackwell B300 chips for around 20 million dollars — a fraction of the cost of comparable closed models.
Why MoE changes the rules for hosting
For a European hosting provider, this kind of mixture-of-experts architecture is significant. In the past, frontier-capable models required enormous hardware to run efficiently. With Trinity Large, you get capacity comparable to a massive model, but with an inference cost that resembles a much smaller one. This makes it realistic to offer advanced AI on owned infrastructure in Europe without blowing the budget.
As Davies Meyer points out, one of the primary advantages of open-weight models is precisely the option of self-hosting. This gives full control over data and makes compliance with rules such as the EU AI Act significantly easier. When a model only activates a fraction of its parameters at runtime, both power consumption and latency drop.
- Lower hardware requirements per request
- Lower inference costs
- Full compliance with European data sovereignty
The road ahead
Arcee AI's Trinity Large is an excellent example of the open-weight movement becoming more practically useful. Models designed to be efficient at runtime, rather than simply impressive on raw parameter count, are exactly what's needed to build a sustainable European AI infrastructure. At Liviate, we keep a close eye on these efficient architectures to make sure our customers can always run the latest AI technology safely and cost-effectively in Denmark.