Guide / TRINITY
Trinity-Large-Thinking: the full numbers, the license mix-up, and where to run it
The complete benchmark table behind Arcee AI's Trinity-Large-Thinking, why the press and the model's own Hugging Face card disagree on its license, and the three real ways to run it — API, download, or OpenRouter.
You commented TRINITY, so here's the rest of it: the full benchmark table (not just the one number), a real discrepancy I found between two of Arcee's own pages, and exactly where to run this thing.
What it actually is
Trinity-Large-Thinking, released by Arcee AI on 1–2 April 2026, is a sparse mixture-of-experts reasoning model:
- ~398 billion total parameters, ~13 billion active per token — a 4-of-256 expert routing strategy (256 experts, 1 shared, ~1.56% sparsity)
- Trained on 17 trillion tokens across 2,048 Nvidia B300 GPUs
- 262,144-token context window on the model card and OpenRouter's listing. Arcee's own marketing separately says "512k" — the two don't agree, and this page uses the card-verified number.
The full benchmark table
The reel said "91.9% on PinchBench, second place, behind only Opus 4.6" — accurate, but it's one row out of the actual comparison table on Trinity's own Hugging Face card. Here's the whole thing, Trinity vs. the four models Arcee chose to compare it against:
| Benchmark | Trinity-Large-Thinking | Opus-4.6 | GLM-5 | MiniMax-M2.7 | Kimi-K2.5 |
|---|---|---|---|---|---|
| IFBench | 52.3 | 53.1 | 72.3 | 75.7 | 70.2 |
| GPQA-Diamond | 76.3 | 89.2 | 81.6 | 86.2 | 86.9 |
| Tau2-Airline | 88.0 | 82.0 | 80.5 | 80.0 | 80.0 |
| Tau2-Telecom | 94.7 | 92.1 | 98.2 | 84.8 | 95.9 |
| PinchBench | 91.9 | 93.3 | 86.4 | 89.8 | 84.8 |
| AIME25 | 96.3 | 99.8 | 93.3 | 80.0 | 96.3 |
| BCFLv4 | 70.1 | 77.0 | 70.8 | 70.6 | 68.3 |
| MMLU-Pro | 83.4 | 89.1 | 85.8 | 80.8 | 87.1 |
| SWE-bench Verified | 63.2 | 75.6 | 72.8 | 75.4 | 70.8 |
Bold = highest in that row. Read plainly: Opus-4.6 wins six of the nine benchmarks shown, GLM-5 wins two, and Trinity wins exactly one outright (Tau2-Airline) — with a strong second-place showing on PinchBench specifically, which is the number Arcee led with and the reel repeated. "Second on PinchBench" is real and sourced. "Second-best model," full stop, is not what this table shows, and I don't want to leave that impression standing.
The license discrepancy
This is worth flagging on its own. Arcee's own blog post says, in plain prose: "Today
we are releasing Trinity-Large-Thinking on our API and the weights on Hugging Face under
the Apache 2.0 license." The Hugging Face model card's actual license field, the one
that governs the repo, reads openmdw-1.1 — Arcee's own "Open Model & Data Weights"
license, not literally Apache 2.0.
I don't know which one is the controlling document. If you're evaluating this for anything where the license terms actually matter, read the license field on the model repo yourself rather than trusting either the press coverage or this page. That's not hedging — it's the one fact here I can't independently resolve for you.
Pricing
$0.90 per million output tokens on Arcee's own API. For comparison, Claude Opus 4.6 is $5 / $25 per million input/output tokens on Anthropic's first-party pricing — so the "96% cheaper" figure reported by press is doing real, checkable math on real numbers, even though it's not a phrase Arcee's own materials use.
Where to actually run it
- Arcee's own API — chat.arcee.ai
- Hugging Face — arcee-ai/Trinity-Large-Thinking, open weights, downloadable
- OpenRouter — listed with reasoning and tool-calling support, no separate setup
Should you actually use it
If your workload is agentic tool-use and cost matters more than shaving the last few points off a benchmark, it's a real, cheap, capable option — genuinely competitive with the open-weight competitors it's compared against here. If you need the best score on any single benchmark in that table other than Tau2-Airline, the table already tells you which model wins it, and it usually isn't this one.
One more thing worth saying plainly: ~398 billion total parameters is not a self-host-on-your-laptop model, even with only 13B active per token. Running it yourself means real infrastructure, not a weekend project. The API and OpenRouter routes exist for exactly that reason.
Related on this site
If you're choosing between Claude models for your own routing setup, see Claude Opus 5 vs. GPT-5.6 model routing.
Sources
- Arcee AI — Trinity-Large-Thinking announcement
- Hugging Face — arcee-ai/Trinity-Large-Thinking model card (primary source for parameters, MoE config, the full benchmark table, and the license field)
- Anthropic's own pricing page, current as of 28 August 2026, for the Opus 4.6 comparison
Verified live 28 August 2026. Product names are factual; no vendor partnership or endorsement is implied.
Published 28 August 2026.