Arcee Trains Frontier Open Models for Long-Running Agents, Powered by NVIDIA Blackwell

Before Trinity-Large-Thinking, the open AI ecosystem had limited options for reasoning models built for long-horizon agents and complex tool use. Trinity-Large-Preview validated market demand for such a model, but its deployment highlighted gaps in multi-turn tool use, context coherence, and instruction following across long-running agent loops.

Developers and enterprises faced a difficult trade-off. Many proprietary models provided strong reasoning, but came with higher serving costs, closed APIs, and limited ability to fine-tune models. Existing open models often struggled to match frontier-level agent performance, especially in workflows with long reasoning chains, numerous tool calls, and durable context retention. Arcee saw a need for open-weight models that organizations could own—inspect, post-train, host, and distill—without being locked into only closed-model economics.

To close this gap, Arcee needed AI infrastructure capable of frontier model training and serving a ~400B parameter sparse mixture-of-experts (MoE) model with ~13B active parameters per token, at high token counts with efficient expert parallelism. This meant finding GPUs and software that could support large-scale pretraining runs, RL post-training pipelines, and production inference for agentic workloads across cloud, on-prem, and edge environments.

Arcee powered its training runs with 2,048 NVIDIA Blackwell Ultra GPUs, leveraging NVIDIA Blackwell Ultra’s high memory capacity, compute density, and NVIDIA NVLink™ and NVLink Switch scale-up fabric to train Trinity Large efficiently. Arcee utilized NVIDIA HGX™ B300’s memory and bandwidth profile—8 TB/s of memory bandwidth per GPU and high-speed NVLink connectivity—to scale expert parallelism efficiently across thousands of GPUs. NVIDIA Quantum InfiniBand networking was used across the Trinity cluster to keep MoE communication overhead low and maintain training efficiency.

The training pipeline incorporated NVIDIA NeMo™ Data Designer for pretraining data preparation and NVIDIA NeMo RL for RL post-training on Trinity-Nano and Trinity-Mini. Synthetic data generation was done using a combination of NVIDIA CUTLASS, CuTe DSL, and cuTile with Arcee’s post-training datasets. Arcee plans to adopt NVIDIA Megatron for large-scale pretraining in the future. 

NVIDIA Dynamo enabled Arcee to manage its vLLM-based serving stack for production inference. To optimize end-to-end performance, Arcee implemented custom NVIDIA® CUDA®, fused MoE kernels with speculative decoding and KV-cache-aware routing to reduce inference latency and lower token costs.

Production inference runs on a mixed fleet with NVIDIA Blackwell Ultra and NVIDIA Hopper-based systems available via Arcee’s API and OpenRouter. The Trinity family of models integrate natively with major agent frameworks and coding-agent harnesses, including OpenClaw, Hermes Agent, and others, for developers to integrate Trinity into their existing production systems.

By training from scratch on NVIDIA Blackwell Ultra GPUs, Arcee validated that open-weight MoE architectures can scale reliably to hundreds of billions of parameters without training loss spikes, while maintaining leading general capabilities. 

Trinity-Large-Thinking has established itself as a frontier-class open reasoning model for long-horizon agents. The model ranks second globally on PinchBench, while leading on select public agentic benchmarks such as τ²-bench Airline. This result provides a strong foundation for future open models and incremental improvements to Arcee’s supervised fine-tuning (SFT) and RL pipelines.

A core outcome of the NVIDIA-powered stack is significantly better economics for frontier reasoning and agent workloads. Trinity-Large-Thinking inference via Arcee’s API is priced at roughly 0.90 USD per million output tokens, which can be around 20x less expensive than closed frontier models based on publicly available pricing. Lowest cost tokens play a critical role in AI factory total cost of ownership (TCO) for long-running agents that generate extensive reasoning traces and multiple tool calls.

Trinity-Large-Preview crossed 3.37 trillion tokens served on OpenRouter in its first two months, and the model became the number-one most-used open model in the U.S. and fourth globally within the OpenClaw collection. In less than three weeks, Trinity-Large-Thinking reached 55.2 billion tokens generated on Kilo Code and 25.6 billion tokens generated on OpenClaw.

Native integrations with OpenClaw and Hermes Agent plus coding agent harnesses from Kilo Code and Open Code reduce integration friction and enable faster prototyping of agentic applications. 

Arcee worked across the broader open model ecosystem, including Prime Intellect for AI compute, Datology for data curation, and OpenRouter and other API partners for developer distribution. These integrations strengthen its position as a frontier U.S. open-weight AI lab, while amplifying reach through APIs, frameworks, and developer tools.