Arcee AI is a small company with only 26 employees, but they just did something that surprised many in the industry: spending about half of their venture capital — $20 million — to run 2,048 Nvidia B300 GPUs for 33 days, they trained an open-source reasoning model with 40 billion parameters called Trinity-Large-Thinking.
The model's agent task performance is quite impressive: it scored first on the Tau2-Airline benchmark (88 points), second on PinchBench (91.9 points, compared to Anthropic's Claude Opus 4.6 at 93.3), and achieved 96.3 on AIME25 math reasoning. It is fully open-sourced under the Apache 2.0 license, allowing commercial use, download, and modification.
Not brute force, but MoE architecture pushed to the limit
Trinity-Large-Thinking is a Mixture of Experts (MoE) model with 40 billion total parameters, but only about 1.3 billion are activated per inference — it uses 256 experts, selecting only 4 each time. This means it has the "knowledge capacity" of a large model while maintaining inference speed and resource usage close to that of a small model.
The training data consists of 17 trillion tokens, of which over 8 trillion are synthetically generated. This ratio is now common among high-quality open-source models.
On the technical side, a highlight is their self-developed SMEBU (Soft-clamped Momentum Expert Bias Updates) algorithm, specifically designed to solve the "expert collapse" problem in large-scale MoE training. The entire 17-trillion-token training process had "zero loss spikes" — a rare achievement in large model training, as many MoE models experience sudden loss explosions midway.
Strengths and gaps
To be honest, Trinity-Large-Thinking does not comprehensively surpass Claude Opus, but excels in specific scenarios.
| Benchmark | Trinity-Large-Thinking | Claude Opus 4.6 |
|---|---|---|
| Tau2-Airline | 88 (1st) | Not disclosed |
| PinchBench | 91.9 (2nd) | 93.3 |
| AIME25 | 96.3 | — |
| GPQA-Diamond | 76.3 | 89.2 |
| MMLU-Pro | 83.4 | 89.1 |
Agent tasks approach top-tier levels, while general reasoning still lags by about 13 percentage points. But considering this is an open-source, locally deployable model, the gap is already quite small.
They also simultaneously released Trinity-Large-TrueBase — a "raw checkpoint" version preserving the pure pre-training state before instruction fine-tuning. This version is valuable for AI fundamental research, allowing study of how RLHF and instruction fine-tuning specifically affect model behavior.
Broader context: The long absence of U.S. open-source AI
Over the past year, the open-source large model leaderboard has been dominated by Chinese companies: DeepSeek, Qwen, Kimi, GLM — new entries every few months. On the U.S. side, Llama has been the sole contender, but Meta recently closed-sourced some new models, drawing criticism from the open-source community. Arcee is filling this gap.
CEO Mark McQuade said the model's target users are enterprises: it can be deployed locally, customized, and does not require sending data to cloud service providers. For financial, healthcare, and government institutions with data compliance requirements, "being able to run on your own server" is sometimes more practical than benchmark scores.
26 people, $20 million, 33 days, 40 billion parameters.
Small companies doing big things — that spirit is still alive in AI.
Sources: Arcee AI spent half its venture capital to build an open reasoning model that rivals Claude Opus in agent tasks (The Decoder); Tiny startup Arcee AI built a 400B-parameter open source LLM from scratch to best Meta's Llama (TechCrunch); CocoLoop, Arcee aims to reboot U.S. open source AI with new Trinity models released under Apache 2.0 (VentureBeat)