
Compile before you deploy.
Muna optimizes LLMs and other AI models before they hit production; shrinking size, boosting performance, and cutting cold starts by up to 45×. Serve the compiled model on Muna, on your own GPUs, or through an OpenAI-compatible endpoint.
Portable
Move the same model from Muna GPUs to your own infrastructure with one CLI command.
Optimized
Compile models into hardware-aware inference servers for the target device.
Compatible
Use the OpenAI SDK, streaming, model aliases, and familiar request shapes.
Any open model.
Every modality.
Your OpenAI client.
Point the official OpenAI SDK at our endpoint — no new concepts to learn. Compiled models serve chat, embeddings, transcription, and speech behind the API you already use.
Large Language Models
Audio and Voice
Embeddings
Tune latency & cost per request.
Serve 3× more.
Compiled models run wherever you point them. Decide where each inference runs at call-time, and prioritize latency, throughput, or cost with extremely fine control.
embedding = muna.beta.openai.embeddings.create(
input="I can choose where each and every inference runs?",
model="@nomic/nomic-embed-text-v1.5",
acceleration="..."
)
No containers.
No cold starts.
Boot 45× faster.
Compilation removes everything between your model and the GPU, so cold starts disappear. The first call lands as fast as the millionth.
From hosted to your compute
in one command.
Start on Muna GPUs, then deploy the same compiled model to Modal, Baseten, or on-prem. Take full ownership of your AI inference stack.
# Deploy a compiled model to your own GPUs
$ muna deploy @openai/gpt-oss-20b \
--provider modal \
--gpu h100

Ready to compile?
Grab an API key and make your first request with the OpenAI client you already have.