Trinity
#54 in Reasoning ModelsArcee · 7× · last seen Jun 30, 2026
Trinity is a family of open Reasoning/Agent language models by Arcee AI, consisting of Trinity Nano (6B), Trinity Mini (26B), and Trinity Large (approximately 400B, Sparse-MoE with ~13B active parameters). The current flagship "Trinity-Large-Thinking" (released April 1, 2026) is a reasoning-optimized model that performs an internal "Thinking" step before generating responses, similar to the earlier Trinity-Mini. The models are provided under Apache 2.0 license with open weights on Hugging Face and are additionally available via Arcee's own API and OpenRouter; key strengths include Tool-Use/Function-Calling, multi-step Agent workflows, and long context windows (up to 262K/512K Token, depending on deployment).
Features
| Key Benchmark (%) | 91.9% on PinchBench (agentic benchmark), ranked #2 behind Claude Opus 4.6 (93.3%) |
| Context Window (Tokens) | 262,144 tokens (Arcee API/OpenRouter serving), architecture supports up to 512K |
| License | Apache 2.0 (open weights, unrestricted commercial use) |
| Multimodality | No, text-only model (no image/audio input) |
| Platform | Arcee API (OpenAI-compatible), OpenRouter, Hugging Face (open weights), self-hosted via vLLM/SGLang/llama.cpp on AWS/GCP/Azure/on-prem |
| Price per 1M Tokens | $0.90 per 1M output tokens (Arcee API); OpenRouter: $0.25 input / $0.80 output per 1M tokens |
| Release Date | April 1, 2026 (Trinity-Large-Thinking); predecessors Trinity Mini/Nano: December 2025 |