Summary
On September 7, 2026, the vLLM project detailed the vLLM TT Plugin, an out-of-tree backend that brings Tenstorrent accelerators to vLLM. Once installed, Tenstorrent hardware auto-registers as a vLLM platform and serves models through the same OpenAI-compatible API, with hand-written TTNN implementations spanning from a two-chip n300 up to a 32-chip Galaxy system using tensor and data parallelism.
What changed
vLLM published the vLLM TT Plugin on September 7, 2026: a self-contained out-of-tree platform plugin that registers Tenstorrent hardware, keeps vLLM's OpenAI-compatible serving surface unchanged, and runs TTNN model implementations across TT mesh systems from n300 to a 32-chip Galaxy.
Why it matters
vLLM's plugin mechanism lets non-NVIDIA accelerators join the dominant open-source serving stack without forking it; a first-class Tenstorrent backend widens hardware choice for LLM inference and pressures the GPU-centric serving market on cost and supply.
Evidence excerpt
vLLM TT Plugin brings Tenstorrent accelerators to vLLM through the standard out-of-tree platform plugin mechanism; the serving surface remains unchanged, using the same OpenAI-compatible API, with hand-written TTNN implementations for TT mesh from a two-chip n300 up to a 32-chip Galaxy.