Trinity-Large-Thinking is an open-weight, reasoning-specialized foundation model released by Arcee AI. Built on a 398-billion parameter sparse Mixture-of-Experts (MoE) architecture activating ~13 billion parameters per token across 256 experts (4 active), it is post-trained with extended chain-of-thought reasoning and agentic reinforcement learning (RL). Designed with native 512K context window support and interleaved thinking-in-context, it targets long-horizon planning, complex multi-turn tool calling, and autonomous software engineering workflows under an open-source license.

Deploying frontier agentic loops often leads to degraded reliability over long sequences due to context forgetting, tool hallucination, and high inference costs. Trinity-Large-Thinking solves these bottlenecks by combining extreme MoE parameter sparsity with native chain-of-thought reasoning traces, delivering frontier-level agent benchmark results with the compute footprint of a much smaller dense model.

Technical Architecture and Model Profile

Dimension / Feature	Trinity-Large-Thinking Specification	Core Developer & Production Benefit
Model Architecture	398B Parameter Sparse MoE (13B Active)	High knowledge capacity with low inference latency and active token cost.
Expert Routing	256 Total Experts (4 Active per Token)	Sigmoid scoring with SMEBU load balancing to prevent routing collapse.
Context Length	Up to 512,000 Tokens (512K)	Ingests full multi-file codebases and extended agent message histories.
Attention Architecture	Hybrid Local Sliding Window + Global GQA	6:1 Grouped Query Attention for maximum concurrent KV cache concurrency.
Reasoning Engine	Native Chain-of-Thought (<thinking> traces)	Interleaved reasoning preserved across multi-step tool-calling loops.
Agent Benchmarks	94.7% τ²-Bench, 91.9% PinchBench, 63.2% SWE-bench	SOTA performance across autonomous agentic and coding suites.
Serving & Deployment	vLLM, SGLang, llama.cpp, OpenRouter, Puter	BF16 and quantized weights (GGUF/AWQ) ready for self-hosting.

Core Breakthroughs & Capabilities

• Thinking-in-Context for Multi-Turn Agent Loops: Generates explicit reasoning traces within <thinking> blocks before emitting function calls or final answers. When these traces are preserved across conversational history, agents avoid cascading logic errors during complex multi-step tasks.

• Sparsity-Driven Inference Economics: By activating only 13B of its 398B parameters per token, Trinity-Large-Thinking keeps per-token generation fast and economical while routing specialized subtasks to dedicated expert groups.

• Out-of-the-Box Agent Framework Compatibility: Engineered specifically for agent runtimes, integrating directly with platforms like OpenClaw, Hermes Agent, and Claude Code without requiring custom prompt wrappers.

• 512K Massive Context Retention: Combines local RoPE positional layers for syntactic accuracy with global NoPE layers to handle long-range reasoning over lengthy documentation sets, system logs, and microservice repos.

• Permissive Open-Weight Availability: Weights and quantized variants are published openly on Hugging Face for enterprise local hosting, on-premise compliance, and custom domain fine-tuning.