What is Trinity-Large-Thinking?
Trinity-Large-Thinking is Arcee AI's flagship, released on 1 April 2026 and still the newest model in its Trinity Large family. It is a mixture-of-experts model with about 398 billion parameters, about 13 billion active per token, trained from scratch in the US and extended to a 512K-token context. It writes a reasoning trace before every answer, and Arcee built it for tool calling and long agent runs; the reasoning must stay in the conversation history for multi-turn and agent use to work.
Arcee pretrained it on 2,048 NVIDIA B300 GPUs and post-trained on 1,152 H100s. It launched under Apache 2.0, and the model card now lists the permissive OpenMDW-1.1 licence. Arcee also publishes the Base and TrueBase checkpoints. On Hugging Face it has 190 likes and 3,171 downloads in the last 30 days.
At launch it was on Arcee's API and OpenRouter, with an output price of $0.90 per million tokens on Arcee. As of 7 November 2026 Arcee's pricing page no longer lists it and OpenRouter shows no active endpoint, so the weights are the way to use it. Arcee says its next Trinity generation is in training.
How much does Trinity-Large-Thinking cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | OpenMDW-1.1 |
| Arcee API (at launch) | $0.90 per 1M output tokens | No longer on Arcee's price list |
Free weights under OpenMDW-1.1. Launch API price $0.90 per 1M output tokens on Arcee; no hosted endpoint listed as of 7 November 2026.
Is Trinity-Large-Thinking worth it?
Trinity-Large-Thinking is a strong tool user and an ordinary reasoner. Arcee's card puts it at 94.7 on Tau2-Telecom, 88.0 on Tau2-Airline and 91.9 on PinchBench, close to Claude Opus 4.6's 93.3 on PinchBench and ahead of GLM-5 and Kimi K2.5 on Airline. On general tests it trails the same rivals: 76.3 on GPQA-Diamond against 89.2 for Opus 4.6, 52.3 on IFBench against 72.3 for GLM-5, and 63.2 on SWE-bench Verified against 72.8.
Those comparisons are seven months old. The models it was measured against have since been replaced by GLM-5.3, Kimi K3 and DeepSeek-V4-Pro, and Nemotron 3 Ultra, a newer US open model, has far more hosting support. Its real distinction is provenance: a large model pretrained from scratch by a US company, with the base checkpoints released too.
The catch: it is now weights-only in practice. Arcee's pricing page lists GLM, DeepSeek and Kimi models but not Trinity, and OpenRouter shows no active provider. Self-hosting about 398 billion parameters in BF16 needs a multi-GPU server, and dropping the reasoning traces from history breaks multi-turn use.
Our take: Trinity is a research and fine-tuning base for teams that want a US-trained model they fully own. For a hosted agent model today, Nemotron 3 Ultra does the same job with providers behind it.
Where it earns its place
A US government contractor with rules on model provenance
Fine-tune a large open model trained by a US company
Start from Trinity-Large-Base or the 10T-token TrueBase checkpoint, both released, and post-train on in-house data. The OpenMDW-1.1 licence allows commercial use and modification.
A customer-support engineering team
Run a self-hosted agent that calls account and billing tools reliably
Serve it with vLLM using the deepseek_r1 reasoning parser and qwen3_coder tool parser, keeping reasoning in the message history. Its 88.0 on Tau2-Airline suits multi-step support flows.
Before you commit
- Setup
- Download from Hugging Face and serve with vLLM in BF16, or use the FP8, W4A16, NVFP4 or GGUF builds Arcee publishes; chat.arcee.ai offers a web chat.
- Integrations
- OpenAI-compatible endpoint through vLLM; works with OpenClaw and Hermes Agent, per Arcee.
- Data and privacy
- Self-hosting keeps data in-house. Arcee's inference API has zero data retention by default, but it no longer lists Trinity; its web app keeps conversation history.
- Cost at scale
- No current per-token price. Self-hosting means serving about 398 billion parameters, a multi-GPU node even with 4-bit builds.
- Limits
- 512K-token context after extension, from an 8,192-token pretraining context; reasoning traces must stay in history for agent loops to work.
What’s good
- 94.7 on Tau2-Telecom and 91.9 on PinchBench, per Arcee
- Permissive OpenMDW-1.1 licence; Base and TrueBase checkpoints also released
- Pretrained from scratch by a US company
- Only about 13B active parameters
What’s not
- No hosted endpoint on Arcee or OpenRouter as of 7 November 2026
- Trails GLM-5 and Opus 4.6 on GPQA, IFBench and SWE-bench in Arcee's own table
- Reasoning traces must be kept in context
- No update since April while rivals moved on
Best for
Teams that want a US-trained open reasoning model to self-host or fine-tune for tool-calling agents.
Trinity-Large-Thinking alternatives
Trinity-Large-Thinking: common questions
- Is Trinity-Large-Thinking worth it?
- A credible American open reasoning model for tool-calling agents, overtaken on most tests and now hard to rent. We rate it 3.4 out of 5.
- How much does Trinity-Large-Thinking cost?
- Free weights under OpenMDW-1.1. Launch API price $0.90 per 1M output tokens on Arcee; no hosted endpoint listed as of 7 November 2026.
- What are the downsides of Trinity-Large-Thinking?
- No hosted endpoint on Arcee or OpenRouter as of 7 November 2026; trails GLM-5 and Opus 4.6 on GPQA, IFBench and SWE-bench in Arcee's own table; reasoning traces must be kept in context; no update since April while rivals moved on.
- Who is Trinity-Large-Thinking best for?
- Teams that want a US-trained open reasoning model to self-host or fine-tune for tool-calling agents.
- How much does Trinity-Large-Thinking cost for a team or in production?
- No current per-token price. Self-hosting means serving about 398 billion parameters, a multi-GPU node even with 4-bit builds.
- Is Trinity-Large-Thinking safe for company data?
- Self-hosting keeps data in-house. Arcee's inference API has zero data retention by default, but it no longer lists Trinity; its web app keeps conversation history.
- What are the best Trinity-Large-Thinking alternatives?
- The closest alternatives we have reviewed are Nemotron 3 Ultra, gpt-oss-120b, GLM-5.3 and DeepSeek-V4-Pro.
Answers checked Nov 7, 2026.