A significant portion of the team behind SGLang, a popular open-source tool used by companies like xAI and Cursor to accelerate AI model training, has moved to its recently launched commercial startup. That company, called RadixArk, was announced last August.
RadixArk originated as the SGLang project in 2023 inside the UC Berkeley lab of Databricks co-founder Ion Stoica. It was recently valued at about $400 million in a funding round led by Accel, according to two people familiar with the matter. The startup had previously raised angel capital from investors, including Intel CEO Lip-Bu Tan.
Ying Sheng, a key contributor to SGLang and a former engineer at xAI, left Elon Musk’s AI startup to become the co-founder and CEO of RadixArk, as announced last month. Sheng was previously a research scientist at Databricks.
Both SGLang and RadixArk focus on optimizing inference processing, which allows AI models to run faster and more efficiently on the same hardware. Together with model training, inference represents a large portion of the server costs for AI services, so tools that optimize the process can create significant savings quickly.
This transition from open-source project to commercial startup mirrors the path of vLLM, a more mature project for optimizing inference. The newly formed vLLM company has had conversations about raising upwards of $160 million at a valuation of about $1 billion. Three people familiar with that potential deal say Andreessen Horowitz is leading the investment, though final numbers remain unconfirmed. vLLM co-founder Simon Mo characterized information about this round as factually inaccurate but declined to specify which details were wrong.
Like SGLang, vLLM was also incubated in Ion Stoica’s lab at UC Berkeley. Several large tech companies already run their inference workloads using vLLM, and SGLang has gained significant popularity over the last six months.
RadixArk continues to develop SGLang as an open-source AI model engine. The startup is also building Miles, a specialized framework designed for reinforcement learning, which allows businesses to train AI models to improve over time. While most of its tools remain free, RadixArk has started charging fees for hosting services.
Startups providing inference infrastructure have seen a surge in funding recently, highlighting the importance of this layer for AI. Baseten recently secured $300 million at a $5 billion valuation. This follows a similar move by rival Fireworks AI, which raised $250 million at a $4 billion valuation last October.