In January 2026, Arcee AI launched Trinity Large Preview. It is a sparse MoE language model with 398–400 billion parameters.
This is the largest open-weight model released by a US laboratory, and the Apache 2.0 license allows anyone to use it for free.
The Trinity Grand Preview is cleverly designed. Even though it has many parameters, it only uses about 13B during inference.
It uses only 1.56% of its parameters, making it fast and efficient compared to dense models.
Trinity AI Family Structure
The Trinity series from Arcee has smaller models (Nano and Mini) and ends with a Large model. In the case of the flagship model, the team publishes three checkpoints from the same training set:
- Trinity-Big Preview got light adjustments to chat and instructions. Use it now for creative talks and assignments. The team continues to improve it with reinforcement learning.
- Trinity Grand Base is a fully trained version, trained on 17 trillion tokens.
- Trinity-Large-TrueBase at 10 trillion tokens with no learning rate changes or instruction adjustments, demonstrating the power of large-scale pure pre-training.
The Preview version is highlighted for offering the best out-of-the-box chat experience.
Architectural Highlights from Trinity-Big-Preview
Trinity-Large-Preview is a rare MoE Transformer. Arcee built it with their own FMoE system called AFMoE.
Key innovations include:
- Trinity-Large-Preview uses 4 of 256 expert routings. Only 4 of 256 experts are active for each token. It provides large capacity but uses little computing.
- Gated attention model (inspired by the latest NeurIPS) to be a better model of long sequences.
- Trinity-Large-Preview combines local and global attention. This keeps performance strong in very long contexts.
- SMEBU is a method that keeps training stable on very large data.
This model officially supports up to 512 thousand context tokens, and in some cases up to 1 million. This makes it suitable for long documents, long conversations, and complex agent tasks.
It was trained on 2,048 GPUs with help from an NVIDIA HGX B200 system
Training costs are approximately $20 million. It takes 30–33 days. This makes it very capital efficient for a frontier model.
Early reviews show Trinity Large Preview competing with top Chinese open models such as DeepSeek, Qwen, and GLM.
This is especially brilliant where most models of pure reasoning fail;
- Literacy, creativity, and prose.
- Consistency of character and role play (no character slander)
- Real-time voice assistance and chat.
- Multiphase agent processes and tool actions.
- Coding related logic and deep code knowledge.
The basic version is already showing strong results. It matches models like GLM-4.5 Based on standard benchmarks.
The Preview version trades some benchmark performance for better chat fluidity and creativity.
Another big advantage is efficiency. Thanks to its sparsity, users earn tokens 2–3× faster than similar models. This is true even with 8-bit quantization.
Why This Release Is Important
Most of the top models remain closed today. Arcee released the 400B scale model openly under the Apache 2.0 license.
It gives developers, researchers, and companies the ability to:
- Refinement/refinement at scale without vendor lock-in.
- Conclude leading-edge classes locally or on cloud hardware at no cost.
- Checkpoint: TrueBase Checkpoint: Raw pre-training dynamics checkpoint.
- Build production AI agents, creative applications, or long-context applications with a truly open model
Trinity Big Preview is free on OpenRouter (during preview) and free on Kilo Code unlimited.
This platform lets you get started immediately without setting up your own resources.
The free preview allowed thousands of people to try a model that typically requires large hardware.
Conclusion from Trinity-Big-Preview
Arcee said Trinity-Large-Preview is still trained with reinforcement learning. Expect better reasoning, better instruction following, and more reliable tools soon.
The team hinted at a future production version. They will focus on complex agents and enterprise tools.
Trinity’s Big Preview is hard to ignore. Anyone who cares about the open source frontier LLM should pay attention.
Small teams can build leading-edge scale models with smart design, good data, and efficiency.
Want to create an agent, write a story, review code, or test a 400B (active 13B) model? Now is the right time. Try it for free at OpenRouter or Kilo Code.
Want to Build AI powered solutions, visit Webkul!
PakarPBN
A Private Blog Network (PBN) is a collection of websites that are controlled by a single individual or organization and used primarily to build backlinks to a “money site” in order to influence its ranking in search engines such as Google. The core idea behind a PBN is based on the importance of backlinks in Google’s ranking algorithm. Since Google views backlinks as signals of authority and trust, some website owners attempt to artificially create these signals through a controlled network of sites.
In a typical PBN setup, the owner acquires expired or aged domains that already have existing authority, backlinks, and history. These domains are rebuilt with new content and hosted separately, often using different IP addresses, hosting providers, themes, and ownership details to make them appear unrelated. Within the content published on these sites, links are strategically placed that point to the main website the owner wants to rank higher. By doing this, the owner attempts to pass link equity (also known as “link juice”) from the PBN sites to the target website.
The purpose of a PBN is to give the impression that the target website is naturally earning links from multiple independent sources. If done effectively, this can temporarily improve keyword rankings, increase organic visibility, and drive more traffic from search results.
