Arcee AI needed to build an open reasoning model for long-running agents that could rival closed frontier systems, but doing so required frontier-scale training, reinforcement-learning post-training, and low-cost inference for a roughly 400B-parameter sparse MoE architecture.
Reported impact
Arcee cut inference pricing to about $0.90 per million output tokens, roughly 20x below leading closed reasoning models, served 3.37 trillion tokens in two months on OpenRouter, and established Trinity-Large-Thinking as a competitive open model for long-horizon agentic workloads.
Arcee trained and served the Trinity model family on a frontier NVIDIA stack built for large open-weight models. Trinity Large was trained on 2,048 NVIDIA Blackwell Ultra GPUs, using HGX B300 systems with NVLink and NVLink Switch fabric plus NVIDIA Quantum InfiniBand to keep expert-parallel MoE training efficient across thousands of GPUs. For data and post-training, Arcee used NVIDIA NeMo Data Designer for pretraining preparation and NVIDIA NeMo RL for reinforcement-learning stages, then combined NVIDIA Dynamo with vLLM, CUDA optimizations, fused MoE kernels, speculative decoding, and KV-cache-aware routing for production inference. This created a repeatable pipeline that spans pretraining, RL, and API serving for Trinity-Large-Thinking, while keeping token costs low enough for long-horizon agents and tool-using workflows where reasoning depth directly affects total cost.
Training data is prepared with NeMo Data Designer, large-scale MoE pretraining runs on Blackwell Ultra HGX clusters over InfiniBand, NeMo RL post-trains Trinity models for agentic reasoning, and Dynamo plus vLLM serve the resulting models through Arcee's API and partner distribution channels.
Arcee combined NVIDIA Blackwell Ultra with NVIDIA Dynamo and vLLM for optimized inference.
Arcee trained Trinity Large on 2,048 NVIDIA Blackwell Ultra GPUs, then post-trained Trinity-Large-Thinking with extended chain-of-thought and agentic reinforcement learning (RL) using NVIDIA NeMo open libraries.
Inference via Arcee's API costs around $0.90 per 1M output tokens, around 20x cheaper than leading closed reasoning models.
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.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Sep 22, 2026 Sluicebox Sluicebox needed to turn fragmented bills of materials, supplier records, PDFs, CAD files, and emissions inventories into timely product carbon assessments, but manual life-cycle-analysis work was slow, expensive, and… 51-80% lower inference costs Lucy · NVIDIA Nemotron 3 Ultra · carbon graph NVIDIA Sample Sep 15, 2026 Lambda Lambda needed to raise token throughput and improve AI-cloud economics, but fixed rack and facility power budgets were leaving GPU capacity stranded and limiting how many nodes could run safely at once. +17% Hybrid Training Throughput Dynamic Power Software · NVIDIA DSX AI Factory Platform · NVIDIA DSX MaxLPS NVIDIA Sample Sep 1, 2026 Onetera Onetera needed to help more than 90 municipalities turn fragmented regulations, zoning maps, permit rules, and legacy documents into fast resident-facing services without imposing heavy IT overhead or expensive manual… 1.7x increase in data throughput GPT OSS 20B · Government Intelligence Agent · LangGraph 