HomeAI models
Arcee AI: Trinity Large Thinking
Arcee AI
•
Trinity Large Thinking
•
Released 
April 2026

Arcee AI
Trinity Large Thinking

Open-weight 398B sparse MoE reasoning model built for long-horizon agents and multi-turn tool calling with preserved thinking.

Modality:
Text
PDF
model ID
arcee-ai/trinity-large-thinking
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

362.38
tok/s

Intelligence Index *

10.8
/ 100

Context Window *

262144
tokens

Input price

1.5
Anytoken

Output price

4.8
Anytoken
Open-Weight Agentic Reasoning at 13B Active Parameters Trinity Large Thinking is Arcee AI's reasoning-optimized variant of Trinity Large. It is a 398B-parameter sparse Mixture-of-Experts model with about 13B active parameters per token. It is post-trained with extended chain-of-thought and agentic reinforcement learning. It sits above the lightly post-trained Trinity Large Preview in Arcee's lineup. It writes explicit reasoning before answers or tool calls. It fits long-horizon agents, multi-turn tool calling, and coding loops where open weights and fast output matter more than top-tier general knowledge. Call Trinity Large Thinking through AnyAPI.ai

Performance

Fast Agent-Loop Execution, Modest General Intelligence

Trinity Large Thinking is tuned for agent loops rather than broad knowledge. Arcee reports 91.9% on PinchBench, second only to Claude Opus 4.6 at 93.3%. Artificial Analysis measures 336 output tokens per second, rated notably fast. The pattern suggests a model that runs tool-driven workflows quickly while trailing frontier models on general intelligence. In production it suits high-iteration agents where output speed compounds, provided reasoning is preserved between turns.

Benchmarks

Artificial Analysis: Fast and Verbose, Below Median on Intelligence

<cite index="12-4">Artificial Analysis gives Trinity Large Thinking an Intelligence Index of 11, below the comparable-model median of 18.</cite> <cite index="12-5">It generated 160M tokens to run the index, somewhat verbose against a 140M median.</cite> Speed is strong at about 336 tokens per second. Reasoning has a visible latency cost: <cite index="11-3">time to first token is 1.26 seconds, but the first answer token arrives at 7.10 seconds.</cite> Treat it as a speed-oriented agent model, not a top-tier general reasoner.

Output Speed

*
362.38
tok/s

Intelligence Index

*
10.8
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
75
%

HLE *

Adherence to multi-step structured instructions.
16
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Trinity Large Thinking takes text input and returns text, including reasoning traces and tool calls. <cite index="22-9">Arcee's model card describes a 512K extended context, while the OpenRouter endpoint exposes 262,144 tokens.</cite> Plan around the limit your provider enforces. <cite index="1-5">Maximum output is listed at 80,000 tokens,</cite> enough for long reasoning chains plus an answer. Reasoning is always on, and thinking must stay in message history for reliable multi-turn behavior.
Verified Specifications — 
Trinity Large Thinking
*
Input modalities
Text
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
262144
Reasoning
Yes
Knowledge cutoff
April 2026
Pricing (standard)
1.5
 AnyTokens in
 / 
4.8
 AnyTokens out

Limitations & Trade-offs

Where Trinity Large Thinking falls short

1
Reasoning delays the first answer. <cite index="11-3">Artificial Analysis measures a 1.26-second time to first token but 7.10 seconds to the first answer token,</cite> because the model reasons before replying, averaging about 35k reasoning tokens per task in its tests. Interactive chat and voice products feel this delay unless the UI streams reasoning. For latency-sensitive responses, choose a non-reasoning model such as Trinity Large Preview.
2
Reasoning must stay in history. <cite index="4-6">Arcee states that thinking tokens must be kept in context for multi-turn conversations and agentic loops to function correctly.</cite> Clients that strip reasoning_content, truncate history aggressively, or use stateless wrappers will degrade tool-calling reliability and inflate context use. Custom orchestration and agent frameworks are most affected. Simple stateless requests are better served by a non-reasoning model.
3
Below frontier on general intelligence and coding. <cite index="12-4">Artificial Analysis scores it 11 on its Intelligence Index, under the comparable-model median of 18.</cite> Its AA-LCR long-context score is 38%, and its AA-Omniscience score is negative, so ground factual answers with retrieval. <cite index="15-3">Arcee-reported SWE-bench Verified is 63.2 against 75.6 for Claude Opus 4.6.</cite> Hard software-engineering and long-document analysis fit higher-tier models better.
4
Hosted limits and no enforced JSON mode. Arcee's model card lists a 512K extended context, but hosted endpoints expose 262,144 tokens, and providers differ on maximum output (80K versus 66K listed). <cite index="20-15">OpenRouter's listing says response_format is not supported, so JSON output is not enforced.</cite> Pipelines needing guaranteed schemas or the full 512K should validate outputs, rely on tool-call arguments, or choose another model.

Best-Fit Workloads

Where this model earns its place

01

Long-horizon tool-calling agents

‍

Arcee developed the model for long-horizon agents, multi-turn tool calling, and context coherence over extended workflows. Arcee reports 94.7% on τ²-Bench, and Artificial Analysis measures 90.1% on τ²-Bench Telecom. It suits planner or executor agents that call many tools per task. Keep reasoning in history, and expect weaker results than top-tier models on the hardest coding steps.

02

Agent framework backends

‍

Arcee's catalog lists compatibility with OpenClaw and Hermes Agent. Kilo reported strong internal results on OpenClaw use cases. At roughly 336 tokens per second, loops with many short iterations finish faster. Best for automation agents that run unattended. Not ideal when each step needs the strongest software-engineering accuracy.

03

Math and analytical reasoning with visible traces

‍

Arcee reports 96.3% on AIME 2025 and 76.3% on GPQA-Diamond, and Artificial Analysis measures 75.2% on GPQA Diamond. Explicit reasoning traces let teams inspect how an answer was reached, which helps debugging and evaluation of analytical pipelines. Factual recall is weaker, so pair it with retrieval for knowledge-heavy questions.

04

Multi-turn service automation


Arcee positions the model for planning, tool calling, and customer-service-style workflows. The 90.1% τ²-Bench Telecom result from Artificial Analysis supports multi-turn, policy-bound tool use. Suited to back-office assistants that look up records and take actions. Its reasoning delay before the first answer token makes it less suitable for real-time voice or chat.

Pricing in anytokens via AnyAPI
Input
1.5
₳
Output
4.8
₳
Cache write
—
₳
Cache read
0.36
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Trinity Large Thinking via AnyAPI.ai

Access Trinity Large Thinking through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Trinity Large Thinking without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Trinity Large Thinking and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Trinity Large Thinking against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Trinity Large Thinking from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Trinity Large Thinking for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Hosted endpoints list a 262,144-token context window for Trinity Large Thinking. <cite index="22-9">Arcee's model card describes a 512K extended context,</cite> so use the limit your provider enforces. <cite index="20-11">Maximum completion length is listed at 80,000 tokens on OpenRouter,</cite> and other providers may list less.

Yes for tool calling, limited for JSON. <cite index="20-14,20-15">OpenRouter lists tools and tool_choice support, but not response_format, so JSON output is not enforced.</cite> Use function-call arguments for structured data, and validate any free-form JSON in your application.

It is strong on agentic tasks but below median on general intelligence. <cite index="13-3">Arcee reports 91.9% on PinchBench, second to Claude Opus 4.6.</cite> <cite index="12-4">Artificial Analysis scores it 11 on its Intelligence Index against a median of 18,</cite> with output near 336 tokens per second.

Yes. <cite index="2-7">Arcee released open weights on Hugging Face alongside its hosted API.</cite> Sources disagree on the license: <cite index="4-1">the Hugging Face card names OpenMDW-1.1,</cite> while press coverage and Arcee's announcement cite Apache 2.0. Check the model card before commercial redistribution.

Yes. <cite index="4-6">Arcee states that thinking tokens must be kept in context for multi-turn conversations and agentic loops to function correctly.</cite> The reasoning is returned in think blocks or a reasoning_content field. Pass it back with each turn, or tool-calling reliability may drop.

* Benchmark data source: Artificial Analysis artificialanalysis.ai