Arcee AI Ships Trinity Large Thinking Under Apache 2.0 — 400B Sparse MoE at $0.85/M Output
Arcee AI has released Trinity Large Thinking, a frontier open reasoning model built on a 400B sparse mixture-of-experts architecture that activates only 13B parameters per token. The weights are published under Apache 2.0. OpenRouter hosts the model at $0.22 per million input tokens and $0.85 per million output tokens — compared to Claude Opus 4.6’s $15/$75.
Architecture
Trinity Large Thinking uses a 4-of-256 expert routing scheme, selecting 1.56% of stored parameters per forward pass. That is the sparsest routing fraction among major frontier MoEs:
| Model | Routing | Fraction |
|---|---|---|
| Trinity Large Thinking | 4-of-256 | 1.56% |
| DeepSeek V3 | 8-of-256 | 3.13% |
| MiniMax-M2 | 8-of-256 | 3.13% |
| Qwen3-235B | 8-of-128 | 6.25% |
| Llama 4 Maverick | 1-of-128 | 0.78% |
Arcee increased expert count from 3 to 6 per token (versus Trinity Large Preview) to stabilise routing at this sparsity level. The model was trained on 17 trillion tokens using 2,048 NVIDIA B300 GPUs with a 256K context window.
Benchmarks
Arcee’s self-reported results position Trinity Large Thinking close to Claude Opus 4.6 on agentic and reasoning tasks, with a significant gap on instruction following:
| Benchmark | Trinity LT | Claude Opus 4.6 |
|---|---|---|
| Tau2-Bench Telecom | 94.7% | 92.1% |
| Tau2-Bench Retail | 88.0% | 82.0% |
| PinchBench | 91.9% | 93.3% |
| GPQA Diamond | 76.3% | 89.2% |
| SWE-bench Verified | 63.2% | 75.6% |
| IFBench | 52.3 | 53.1 |
| AIME 2025 | 96.3% | 99.8% |
The Tau2 numbers — which test multi-turn tool-using agent capability in retail and telecom environments — beat Opus 4.6 on both dimensions. SWE-bench at 63.2% is the weakest result: competitive with mid-tier open models, not frontier-class coding performance. Results are self-reported and have not yet been independently replicated.
Pricing and cost structure
At $0.85/M output, Trinity Large Thinking is 98.9% cheaper than Claude Opus 4.6 ($75/M output) on a per-token basis. For output-heavy agentic pipelines running hundreds of thousands of agent turns, the cost difference is structural. A workload spending $10,000/month on Opus 4.6 output would cost roughly $113/month on Trinity Large Thinking — assuming comparable task success rates, which the SWE-bench gap suggests is not a given for complex coding tasks.
Distribution
Weights are available on Hugging Face under Apache 2.0. OpenRouter carries the model at pay-as-you-go rates; DigitalOcean’s Agentic Inference Cloud hosts a managed preview endpoint. The base instruct model (Trinity Large Preview) was renamed on Arena on April 10 to clarify it is still in preview — the Thinking variant represents Arcee’s production reasoning release.
What it means
Trinity Large Thinking is credible on agentic multi-turn benchmarks and priced for volume. The SWE-bench gap is real: 63.2% versus frontier scores above 80% means it is not a drop-in replacement for coding-heavy agent workflows. Where it fits is cost-sensitive agentic chains where telecom-style tool use matters more than raw code generation — customer service automation, structured data extraction, long-horizon planning tasks. The fully open Apache 2.0 licence also makes it the only frontier-class reasoning model enterprises can self-host without negotiating commercial terms.