Trinity Models
Reliable intelligence you can run anywhere.
Open weights and a production API for multi-turn conversations, tool use, and structured outputs. Pick your size based on where you deploy. Capabilities stay consistent.
Models
All Trinity variants share the same skill profile. Choose the footprint that fits your infrastructure.
Capabilities
- Agent reliability: accurate function selection, valid parameters, schema-true JSON, and graceful recovery when tools fail.
- Coherent multi-turn conversation: holds goals and constraints over long sessions and follows up naturally without re-explaining context.
- Structured outputs: JSON schema adherence with native function calling and tool orchestration.
- Same skills across sizes: move workloads between edge and cloud without rebuilding prompts or playbooks.
- Super efficient attention: reduced cost of running at long contexts compared to other models.
- Strong context utilization: makes full use of large input docs for more relevant, grounded responses.
Technical Overview
Architecture
Sparse mixture of experts with highly efficient attention for lower latency and lower cost.
- Trinity Mini: 26B total / 3B active per token
- Trinity Nano: 6B total / 1B active per token
- Trinity Large: 400B total / 13B active per token
Training
- Curated, high-quality data across diverse domains with strict filtering and classification pipelines.
- Heavy synthetic augmentation for edge cases, tool calling, schema adherence, error recovery, preference following, and voice-friendly styles.
- Evaluation focused on tool reliability, long-turn coherence, and structured-output accuracy.
Context & I/O
- 256K token context window
- Diverse MCP support
- Parallel tool use and raw reasoning traces
Deploy Your Way
Quickstart
- Get an API key or download the weights.
- Pick Mini for cloud or Nano for edge.
- Choose Reasoning or Standard per workload.
- Plug in your tools.
- Ship to production.
FAQs
Trinity is a family of open-weight language models built for multi-turn conversations, tool use, and structured outputs. Run it fully locally or use a hosted API — the capabilities stay consistent across sizes.
All models share the same capabilities, APIs, and skill profile, so you can move between them without changing prompts.
- Trinity Nano (6B) is optimized for offline operation and low-latency loops on embedded or edge devices.
- Trinity Mini (26B) is tuned for multi-turn agents, tool orchestration, and structured outputs in cloud or on-prem backends.
- Trinity Large (400B) excels at agentic loops, context coherence, and instruction following across long-horizon runs.
It selects the right tool for each task, produces valid parameters and schema-compliant JSON, recovers gracefully when a tool fails, and maintains coherence over 10–20 turns — consistently across model sizes.
A sparse Mixture-of-Experts architecture (only a subset of experts activate per token) trained on curated, domain-diverse data with synthetic augmentation for tool calling, schema adherence, and error recovery.
Nano and Mini support a 128K-token context window, with Large pushing up to 256K — enough for long conversations, multi-step workflows, and structured outputs without losing coherence.
Yes. Trinity can serve as a streaming LLM backbone for voice workflows, feeding real-time output into any text-to-speech system for low-latency, multi-turn interactions.
Yes. Arcee provides an OpenAI-compatible API endpoint, making it easy to integrate into existing systems.
Yes. Trinity natively supports structured outputs, including JSON and other schema-based formats — define your schemas and the model adheres to them.
- Try it on the platform to explore the models immediately.
- Generate an API key to integrate Trinity into your own apps.
- Download the open weights from Hugging Face to run anywhere.