RadixArk Raises $100M Seed Funding for AI Tools

RadixArk Inc., an emerging player in the artificial intelligence development ecosystem, has successfully closed a substantial $100 million seed funding round. This significant capital infusion was spearheaded by Nvidia Corp.'s venture arm, NVentures, alongside prominent venture capital firm Spark Capital. The round also saw participation from industry heavyweights including Advanced Micro Devices Inc., Databricks Inc., and Broadcom Inc. Chief Executive Hock Tan, underscoring strong confidence in RadixArk's technological vision. The company's valuation now stands at an impressive $400 million following this investment.

The startup is focused on commercializing advanced AI development tools derived from two key open-source projects: SGLang and Miles. The team behind RadixArk played a pivotal role in the initial development of SGLang before the company's inception. Miles, a project that RadixArk open-sourced in November, is designed to significantly accelerate reinforcement learning, a critical training methodology for large language models (LLMs). Miles offers the capability to compress LLMs with up to one trillion parameters, enabling them to fit within the memory of a single high-end graphics processing unit, thereby drastically reducing computational costs.

Further enhancing AI project efficiency, Miles incorporates an innovative asynchronous co-evolutionary framework named MrlX. This framework facilitates the simultaneous training of multiple AI agents within a shared simulated environment. Such an arrangement allows agents to refine their reasoning abilities through mutual learning and interaction, a crucial step in developing more sophisticated AI systems. The ability to train complex models more affordably and efficiently addresses a major bottleneck in the current AI development cycle, a sector projected to grow exponentially in the coming years.

Post-training, developers can leverage SGLang for AI inference, the process of using a trained model to generate predictions or outputs. SGLang provides essential building blocks for constructing robust inference environments. According to company statements, SGLang is already powering AI clusters aggregating over 400,000 graphics cards, demonstrating its scalability and performance in demanding applications. The demand for efficient inference solutions is escalating as more organizations deploy AI models into production.

A key innovation within SGLang is its ability to optimize the KV cache, a data structure crucial for LLM prompt processing. By enabling LLMs to reuse KV cache data across successive prompts, SGLang circumvents the need to regenerate this temporary data from scratch for every request. This optimization significantly lowers infrastructure overhead and accelerates response times, a critical factor for user-facing AI applications. The AI inference market is expected to reach hundreds of billions of dollars globally within the next decade, driven by the widespread adoption of generative AI.

In addition to KV cache reuse, SGLang employs speculative decoding to offload tasks to less resource-intensive models and supports distributed computation across heterogeneous hardware architectures. These advancements collectively boost model performance and flexibility. RadixArk plans to offer managed infrastructure and tooling for cloud-based AI model hosting, alongside continued enhancements to its open-source projects, aiming to capture a significant share of the rapidly expanding AI infrastructure market.