Triton Inference Server

Triton Inference Server is NVIDIA's open-source inference serving software, and it has a naming wrinkle. An NVIDIA developer post says that as of March 18, 2025 it became part of the NVIDIA Dynamo platform and was renamed NVIDIA Dynamo-Triton, and NVIDIA's Dynamo-Triton page describes it as formerly Triton Inference Server. NVIDIA's Dynamo page calls Dynamo the successor to Triton, though third-party blogs go further and say Triton is being replaced, which is interpretation rather than NVIDIA's wording. One aggregator says both names remain in use in 2026 and that the code and containers still carry the Triton name, so the open-source project and container names have not changed.

Triton serves models from several frameworks and has backends for vLLM and TensorRT-LLM, according to a 2026 deep dive. The software is free to download. My searches did not cover the licence terms or any NVIDIA AI Enterprise pricing, so check the GitHub LICENSE file and NVIDIA's AI Enterprise licensing page. One source notes that NVIDIA NIM containers are distributed under NVIDIA AI Enterprise, which is a separate paid track, but gives no price. You also pay for the GPU or CPU hardware, and Triton can run on both.

Triton suits platform teams that need one server for many model types with batching and multi-GPU support. Compare KServe, TensorRT-LLM and BentoML. Browse the MLOps & Infrastructure tools directory for more.

Multi-framework model serving

GPU and CPU.

LLM serving via backends

vLLM, TensorRT-LLM.

Was Triton renamed?

Dynamo-Triton since March 2025.

Is it free?

The software is; AI Enterprise is separate.

Is Dynamo replacing it?

NVIDIA calls Dynamo the successor.

Runs on CPU?

It can run on CPU and GPU.