News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnTrinity Large

Trinity Large

398B-parameter sparse MoE model family by Arcee AI

Arcee AI released Trinity Large, a family of large language models with 398 billion total parameters. The system uses a sparse mixture of experts design to keep active computation low while supporting long contexts and complex reasoning tasks.

Share on WhatsApp

  • 398 billion total parameters with 13 billion active
  • Supports context windows up to 512,000 tokens
  • Trained on 17 trillion tokens using NVIDIA B300 GPUs
  • Generates explicit reasoning traces for agentic tasks
  • Released under OpenMDW-1.1 licence

What it is

Trinity Large is a family of language models built by Arcee AI. The largest variant holds 398 billion parameters in total, though only about 13 billion are active for any given token due to its sparse design. This architecture allows the system to handle very long inputs, up to 512,000 tokens, while maintaining efficiency. The family includes base models for general use and specialised variants optimised for reasoning and agent workflows. It supports text generation and structured tool calling, making it suitable for complex automated tasks.

How it came about

Arcee AI developed the Trinity family using data from partner Datology. The training process involved pretraining on 17 trillion tokens, with an intermediate checkpoint saved after 10 trillion tokens. The team used 2,048 NVIDIA B300 graphics processing units for computation, supported by infrastructure partner Prime Intellect. The family was released with multiple checkpoints, including a base model, a lightly post-trained preview version, and a thinking-optimised variant. A technical report detailing the training methods was published alongside the release.

Inside the model

The model uses a sparse mixture of experts architecture, specifically the AfmoeForCausalLM design. It routes each token to four out of 256 possible experts, resulting in high sparsity. Pretraining began with an 8,192 token context window, which was later extended to 512,000 tokens. Post-training involved instruction tuning on 20 billion tokens. The thinking variant underwent additional reinforcement learning focused on agentic tasks and chain-of-thought reasoning. This variant generates explicit reasoning traces enclosed in specific tags before providing final answers.

What it can do

The thinking variant achieves high scores on agentic benchmarks, including 94.7 per cent on Tau2-Bench and 98.2 per cent on LiveCodeBench. It performs strongly on coding and mathematics tests, such as scoring 96.3 on AIME25. General knowledge benchmarks show competitive results, with 83.4 on MMLU-Pro. The model is designed to preserve reasoning context across multiple turns, which is critical for multi-step planning. Without these reasoning traces, performance on complex agent tasks may degrade, leading to malformed tool calls.

How to use it

Weights are available under the OpenMDW-1.1 licence, allowing for local deployment. The model runs on frameworks such as vLLM, Transformers, and llama.cpp. Arcee AI also provides API access through OpenRouter, where both the preview and thinking variants are hosted. The thinking variant requires specific parser settings in vLLM to separate reasoning content from final responses. Developers are advised to keep full reasoning history in chat loops to maintain performance.

Questions people ask about Trinity Large

How many parameters does Trinity Large have?

The model has 398 billion total parameters. However, it uses a sparse mixture of experts design, so only about 13 billion parameters are active for each token processed. This allows for high capacity with lower computational cost per step.

What is the difference between the Preview and Thinking variants?

Trinity-Large-Preview is a lightly post-trained chat model. Trinity-Large-Thinking includes additional training for reasoning and agentic workflows. It generates explicit reasoning traces in special blocks before giving its final answer, which helps with complex multi-step tasks.

Can I run Trinity Large locally?

Yes, the weights are open under the OpenMDW-1.1 licence. You can run it using tools like vLLM, Transformers, or llama.cpp. Ensure your hardware can support the model size, or use quantised versions available for llama.cpp and LM Studio.

Why is preserving reasoning content important?

For the Thinking variant, the reasoning traces contain the model's internal logic. If these are removed from the chat history during multi-turn conversations, the model may lose context. This can lead to errors in tool calling or poor performance on complex agentic loops.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Trinity Large yet, so the facts on this page come from the sources listed above: huggingface.co/arcee-ai/Trinity-Large-Thinking, huggingface.co/arcee-ai/Trinity-Large-Preview. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.