Trinity
Trinity is Arcee’s Apache-2.0 MoE family. OpenTryOn calls the first-party OpenAI-compatible Platform API for hosted ids. The models are text in, text out.
| Registry id | Service | Upstream id | Auth |
|---|---|---|---|
trinity-large | understand | Trinity-Large-Preview | ARCEE_API_KEY |
trinity-large-thinking | understand | Trinity-Large-Thinking | ARCEE_API_KEY |
trinity-mini | understand | trinity-mini | ARCEE_API_KEY |
trinity-nano | understand | Trinity-Nano | local server (TRINITY_BASE_URL) |
Trinity Large Preview is 400B / 13B-active, hosted at 128K context (weights support 512K). Trinity Large Thinking is the reasoning-optimised Large; responses may include reasoning_content. Trinity Mini is 26B / 3B-active. Trinity Nano (6B / 1B-active) is not hosted on the Arcee Platform — point --model trinity-nano at your own vLLM/SGLang server.
Base URL is https://api.arcee.ai/api/v1. Override with ARCEE_BASE_URL or --base-url. ARCEEAI_API_KEY is an alias for ARCEE_API_KEY. Nano uses TRINITY_BASE_URL (default http://127.0.0.1:8000/v1).
opentryon understand --model trinity-large \
--prompt "Write a 3-sentence lookbook caption for a linen trench."
opentryon understand --model trinity-large-thinking \
--prompt "Plan a three-look campaign around one linen trench." \
--system "You are a fashion editor."
opentryon understand --model trinity-nano \
--prompt "Tighten this PDP copy."
Images and video are rejected. Use kimi-k2.6, moondream-3.1, or nemotron-omni for media.
Large self-host needs multi-GPU; there is no in-process GPU registry id.
MCP tools: understand_trinity_large, understand_trinity_large_thinking, understand_trinity_mini, understand_trinity_nano. Named-model-only — the understand default stays kimi-k2.6.