Trinity Large Thinking
#36 in Reasoning-Modellearcee · large thinking · siet 2026-04-01 · 12× · tolest 30. Juni 2026
5
Momentum
Trinity-Large-Thinking is a reasoning model by Arcee AI with 398 billion total parameters in a sparse Mixture-of-Experts architecture (256 experts, 4 active per token, ~13B active parameters). It was pretrained on 17 trillion tokens and post-trained with extended chain-of-thought and agentic reinforcement learning. The model generates explicit reasoning traces inside <think>…</think> blocks and is specifically designed for multi-step agent workflows, tool calling, and structured outputs. Weights are freely available on Hugging Face; initially released under Apache 2.0, the license was subsequently updated to OpenMDW-1.1 (Linux Foundation).
Momentum-Verloop
04.04.03.07.
Features
| Key Benchmark (%) | τ²-Bench: 94.7% | PinchBench: 91.9% (#2 behind Claude Opus 4.6) | LiveCodeBench: 98.2% | SWE-bench Verified: 63.2% | MMLU-Pro: 83.4% (all values per official Arcee documentation) |
| Context Window (Tokens) | 262,144 tokens (via API/OpenRouter); official docs also mention 512K tokens for extended-context agentic loops |
| License | Originally Apache 2.0 (at launch); later switched to OpenMDW-1.1 (Linux Foundation, permissive, for AI model distributions) – no usage restrictions, commercial use allowed |
| Multimodality | Text-only; no image or audio modality |
| Platform | Arcee AI API, OpenRouter, Hugging Face (open weights, 797 GB); self-hosting via vLLM, SGLang, llama.cpp, LM Studio, Transformers; cloud deployment on AWS, GCP, Azure, on-prem |
| Price | Weights free on Hugging Face (open weights); API usage: $0.25/1M input, $0.80/1M output (Arcee platform, per official pricing page) |
| Price per 1M Tokens | Arcee API (official): $0.25/1M input tokens, $0.80/1M output tokens; OpenRouter: $0.25 input / $0.80 output |
| Release Date | April 1, 2026 |