← All organizations

AI model compilation and inference infrastructure

Muna

Muna compiles AI models into hardware-aware inference software for developers building AI applications. Its platform supports language, vision, audio and embedding models, with an OpenAI-compatible API for chat, embeddings, transcription and speech. Developers can serve compiled models on Muna’s GPUs, deploy them on their own infrastructure or run them directly on users’ devices. The platform lets developers choose where inference runs and prioritize latency, throughput or cost.

Founder Yusuf Olokoba developed Muna’s compiler approach after encountering difficulties distributing computer-vision models. Its compiler traces Python functions into an intermediate representation, propagates type information and generates C++ for compilation across platforms. Muna uses AI-generated operator implementations and measures hardware-specific variants to select faster execution paths. This is ahead-of-time compilation: the approach aims to distribute model executables without bundling a Python interpreter, rather than requiring developers to rewrite their inference code in C++.

Muna charges for hosted compute by the second and for on-device model deployments, with subsequent local predictions free of additional inference charges. Team and enterprise offerings add capabilities including usage analytics, dedicated GPU capacity and on-premises or VPC deployment. In 2025, the company reported serving thousands of unique devices each month with compiled Python functions.

www.muna.ai

1 talk

Newest first

1 speaker at AIE

Affiliations reflect their AIE appearances, not necessarily current employment.

Messages from the stage

Beyond PyTorch tracing

Olokoba explains why PyTorch tracing was insufficient for arbitrary Python dependencies and describes using LLMs to generate C++ and Rust implementations of compiler operations.

Affiliations reflect each recorded session, not necessarily current employment.

Company sources · checked 2026-08-28