Goethe

A company with its own NVIDIA GPUs can point existing code at Goethe instead of a hosted API. The project's example keeps the OpenAI Python SDK and changes only the address, to a /v1 endpoint on port 8443 of the customer's machine. The README also says clients for Anthropic and LiteLLM work, and lists LangChain, LlamaIndex and Open WebUI.

## What is in the repository

There is a goethe command, a background service, a gateway and an installer, plus GPU kernels for NVFP4 and ternary models. After install, goethe start walks through choosing a model, serving it and creating API keys.

## Hardware limits

It needs Linux on x86_64 or ARM64 with glibc 2.30 or newer, and NVIDIA GPUs from Ampere through Blackwell, which rules out Volta and Turing. Driver 580 or newer covers every listed card, drivers 535 to 579 cover only data centre GPUs, and anything older is unsupported. No separate CUDA toolkit is required.

## Key facts

**Maker:** meshapplied **Launched:** 7 November 2026 (Show HN) **Licence:** Apache-2.0 **Price:** free to run; no paid plan seen **Website:** meshapplied.com

Not yet independently verified: Dated from the makers' Show HN post of 7 November 2026; the repository lists no releases. Speed and compatibility claims are the makers' own, with no benchmark on the page, and no company address or country is given. We will update this when it can be confirmed, and remove this note.