RadixArk

RadixArk is a deep-tech company focused on developing large-scale infrastructure systems for AI training and inference. The firm spun out of SGLang, a popular open-source project from a UC Berkeley lab that helps accelerate the inference stage of AI models to reduce computing costs and improve performance. RadixArk was co-founded by CEO Ying Sheng, a former engineer at xAI and research scientist at Databricks. The company recently achieved a valuation of approximately $400 million, with funding from backers including venture capital firm Accel and angel investors such as Intel CEO Lip-Bu Tan.

The company's mission is to make frontier-level AI infrastructure open and accessible to the entire AI community, which includes individual developers, research labs, startups, and large enterprises. RadixArk operates on an open-core business model. It continues to develop and support the open-source SGLang engine for model inference while building proprietary commercial offerings. Its product suite includes managed infrastructure and production-grade tooling for model inference and post-training. A key offering is 'Miles,' an open-source framework designed for large-scale reinforcement learning (RL) and post-training. Revenue is generated through managed hosting services and other commercial products built on top of its open-source foundations.

RadixArk aims to solve the problem of duplicated effort and siloed knowledge in AI infrastructure, where many companies rebuild similar systems from scratch. By providing robust, open, and scalable systems, the company enables the AI community to train, deploy, and operate advanced models more efficiently and affordably. The team consists of engineers with experience shipping production AI systems at companies like xAI and contributing to infrastructure that powers leading AI labs.

Keywords: AI infrastructure, large-scale inference, model training systems, deep-tech, SGLang, reinforcement learning framework, open-source AI, Miles framework, AI model deployment, AI inference optimization, production AI, GPU infrastructure, TPU systems, compute infrastructure, model post-training