RadixArk
26 known investors
RadixArk builds open-source AI infrastructure for large-scale model training and inference, including the SGLang inference engine and Miles reinforcement learning framework. The company provides managed infrastructure and tooling for AI builders ranging from individual developers to enterprises and research labs.
Also known as Project SGLang Β· SGLang
Founders & leadership
Investors Β· 26
Also in the syndicate Β· 18
Funding
SEC filings, press & company announcements$100M disclosed across 1 of 3 rounds Β· 2026
- $100MSeed FundingMay 2026 Β· 5 sources
Accel (lead), Spark Capital (lead), A&E Investment, AMD, Databricks, Eric Zelikman, Hock Tan, HOF Capital, Igor Babuschkin, John Schulman, LDV Partners, LDVP, Lip-Bu Tan, Logan Kilpatrick, MediaTek, NVentures, NVIDIA's NVentures, Olivier Pomel, Robert Nishihara, Salience Capital, Soumith Chintala, Thomas Wolf, Walden Catalyst, Walden Catalyst Ventures, William Fedus, WTT Fubon Family
Source β
Source: company announcements and press reports β follow each round's link for the claim.
Valuation Β· disclosed
Disclosed eventsSource: SEC prospectus filings, and round valuations the company or its investors disclosed β follow each entry's link for the claim.
Company profile
researched Aug 2026RadixArk is a Palo Alto, California-based AI infrastructure company building large-scale inference and training systems. It is the commercial entity behind SGLang, an open-source inference engine, and Miles, an open-source framework for large-scale reinforcement learning and post-training. On top of these two open-source cores, the company ships managed infrastructure and tooling intended for individual developers, startups, enterprises and research labs.
The company frames itself as "infrastructure-first," arguing that advanced training and inference stacks currently sit inside a small number of large labs and that every new AI team rebuilds the same schedulers, compilers, serving engines, orchestration layers, evaluation pipelines and training infrastructure. Its stated aim is to make frontier-level infrastructure open and broadly accessible, with a target of making building, training and running frontier models at least 10x cheaper and 10x more accessible.
SGLang, created in 2023, is described by the company as a de facto open-source standard for inference, serving trillions of tokens daily for organizations including Google, Microsoft, NVIDIA, Oracle, AMD, LinkedIn, xAI, Thinking Machines Lab and humans& (an investor post also lists Nebius; a third-party profile also lists Cursor/Anysphere). Miles, introduced more recently, targets large-scale post-training and has reportedly been adopted by industry teams for large-scale Mixture-of-Experts training.
Founding story
SGLang was created in 2023 as an academic research project; one account places its origin within LMSYS, a nonprofit research collective spanning Stanford, UC Berkeley, Carnegie Mellon and UC San Diego, and another within the UC Berkeley lab of Ion Stoica. Ying Sheng, a primary SGLang contributor with a Stanford PhD in computer science focused on high-throughput generative inference, later built inference systems for Grok at xAI and previously worked as a research scientist at Databricks; she left xAI to become co-founder and CEO. She co-founded RadixArk with Banghua Zhu, described in one source as having worked on systems at NVIDIA. One source states the two founded the company in 2025; investors reported meeting the founders before the company was formally incorporated.
Business model
RadixArk follows an open-core model: the SGLang inference engine and the Miles RL/post-training framework are open source and free to use, while the company sells managed infrastructure and tooling layered on top. Sources describe a commercial platform that productizes the operational complexity of running these engines at production scale β managed inference, managed training and fine-tuning services β positioned for organizations that want to own and operate their own models rather than rent tokens through third-party APIs. A person familiar with the company told TechCrunch that while most of RadixArk's tools remain free, it has begun charging fees for hosting services.
Paid managed services: hosting/managed inference and managed training and fine-tuning offerings sold on top of the freely available open-source SGLang and Miles projects. TechCrunch reported that most tools remain free while the company has started charging fees for hosting services.
Traction
SGLang is described by the company and investors as a de facto open-source standard for inference since 2023, powering trillions of tokens daily and running on hundreds of thousands of GPUs, with named production users among major technology companies and AI labs. Miles, introduced more recently, has reportedly been adopted by industry teams for large-scale MoE training. Commercially, the company had begun charging for hosting services as of early 2026 and raised $100 million at a $400 million post-money valuation; no revenue figures are disclosed. One profile lists team size as 1-10.
Latest developments
In January 2026, TechCrunch reported that RadixArk had been valued at roughly $400 million in an Accel-led round and that SGLang maintainers had moved to the company; it also reported prior angel funding including from Intel CEO Lip-Bu Tan and that RadixArk had begun charging for hosting services. In May 2026, the company publicly announced its launch with $100 million in seed funding at a $400 million post-money valuation. A third-party profile dates the Miles RL framework launch to July 2026 and notes expansion beyond core inference optimization into reinforcement learning for Mixture-of-Experts models.
βΈFull profile β market position, technology, go-to-market, geography, history, risks & controversies
Market position
Positioned as a high-performance AI inference and training infrastructure provider for AI labs, enterprises and high-throughput LLM application developers. Coverage places RadixArk within a wave of open-source AI infrastructure projects converting into venture-backed companies, alongside vLLM (also incubated in Ion Stoica's UC Berkeley lab), and in a funding environment that has recently included Baseten ($300M at a $5B valuation) and Fireworks AI ($250M at a $4B valuation). A CRV general partner told TechCrunch that SGLang gained significant popularity over the prior six months; one profile describes it as a primary alternative to vLLM for high-performance inference.
Third-party profile material attributes RadixArk's technical differentiation to RadixAttention, a token-level radix-tree prefix caching technique enabling automatic KV cache sharing across requests, contrasted with vLLM's block-level hashing, and to a compiler/runtime for structured generation in multi-step and multi-turn LLM workflows. Investors emphasize a distribution advantage: SGLang's existing production adoption at major AI organizations, developer trust from open-source origins, and a community already building on the framework, which the company intends to convert into an enterprise infrastructure platform. The company itself emphasizes first-principles systems design, open publication of code, benchmarks and architectural insights, and reliability at frontier scale.
Technology
The stack centers on SGLang, an open-source inference engine, and Miles, an open-source reinforcement learning and post-training framework, plus managed infrastructure and tooling built above them. Third-party analysis describes SGLang's core mechanism as RadixAttention: a radix-tree data structure that caches previously computed prompt prefixes so overlapping context is reused rather than recomputed, reducing redundant computation, per-token cost and latency β the source of the "Radix" in the company name. The framework also supports structured generation and multi-step LLM workflows via a specialized compiler and runtime, and the company states it will extend support to new model architectures and hardware platforms.
Go-to-market
Bottom-up, open-source-led distribution: SGLang and Miles are released freely and have been adopted organically by large AI operators, which the company and its investors describe as a flywheel into a commercial managed platform. Sources note SGLang achieved adoption without a marketing or sales organization. Monetization is layered on via paid hosting and managed services, with recruiting and community engagement conducted through the company's website, blog and X account.
AI labs, enterprises, cloud and software companies, startups, research labs and individual developers running high-throughput LLM inference or large-scale post-training. Named SGLang users include Google, Microsoft, NVIDIA, Oracle, AMD, LinkedIn, xAI, Thinking Machines Lab, humans&, Nebius and Cursor.
Geography
Headquartered in Palo Alto, California; one profile describes a North America base (Berkeley/San Francisco area) with a global open-source developer community.
History
SGLang originated in 2023 as a research project; TechCrunch describes it as beginning inside the UC Berkeley lab of Databricks co-founder Ion Stoica, while other sources attribute its incubation to LMSYS, a nonprofit research collective founded by students from Stanford, UC Berkeley, Carnegie Mellon and other institutions. A profile source lists an open-source launch of SGLang at UC Berkeley SkyLab and a commercial spin-out as RadixArk in 2025. RadixArk was publicly launched in late 2025 per an investor account, with TechCrunch noting in January 2026 that part of the SGLang maintainer team had transitioned to the new company and that it had been valued at about $400 million in an Accel-led round. The company formally announced its launch and $100 million seed round in May 2026, and introduced the Miles RL framework, dated July 2026 by one profile source. Reported dates across sources are inconsistent.
Risks & controversies
Key facts are unevenly sourced: TechCrunch reported the January 2026 valuation via unnamed sources and could not confirm round size, and RadixArk, Accel and Lip-Bu Tan did not respond to requests for comment. Sources also conflict on founding and launch chronology (2023 project origin; 2025 incorporation/public launch per some sources; May 2026 launch announcement) and on the SGLang incubation venue (LMSYS versus Ion Stoica's UC Berkeley lab/SkyLab). Investor lists vary slightly across sources in naming individuals and entities. The company competes directly with better-established vLLM, which was reported to be raising up to $160 million at roughly a $1 billion valuation, and with large incumbents including NVIDIA's own serving stack, while several of its strategic investors (NVIDIA, AMD, MediaTek, Databricks) are also ecosystem participants. No registered patents are disclosed.
Compiled by commissioned research from 8 cited public sources β announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed β not independently audited.
Competitors Β· 2
by search overlapCompanies competing with RadixArk for the same Google search keywords, organic and paid, via search-intersection analysis.
Timeline Β· 5
launches, deals, and filingsRadixArk introduced Miles, an open-source framework for large-scale reinforcement learning and post-training, reported as adopted by industry teams for large-scale Mixture-of-Experts training.
RadixArk announced its official launch alongside a $100 million seed round at a $400 million post-money valuation, led by Accel and co-led by Spark Capital, with participation from NVentures, Spark, AMD, MediaTek, Databricks and numerous angels. Proceeds are earmarked for growing SGLang and Miles, supporting new model architectures and hardware platforms, and building managed infrastructure.
$100M source β
Prior to the seed round, RadixArk reportedly raised angel capital from investors including Intel CEO Lip-Bu Tan.
TechCrunch reported, citing two people familiar with the matter, that RadixArk β the commercial company behind the SGLang open-source inference engine β was recently valued at about $400 million in a round led by Accel, and that part of the SGLang maintainer team had moved to the startup. TechCrunch could not confirm the round size.
Ying Sheng, a key SGLang contributor, former xAI engineer and former Databricks research scientist, announced on LinkedIn that she had left xAI to become co-founder and CEO of RadixArk.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
βΈResearch sources Β· 8
primary sources listed
- RadixArkradixark.ai Β· web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does RadixArk do?
- Palo Alto AI infrastructure startup commercializing the open-source SGLang inference engine and Miles RL training framework.
- Who founded RadixArk?
- RadixArk was founded by Banghua Zhu.
- Who are RadixArk's investors?
- RadixArk's investors include Accel, HOF Capital, Vocal Ventures, LDV Partners, MediaTek, NVentures, Spark Capital, Walden Catalyst.
- How much funding has RadixArk raised?
- RadixArk has disclosed $100M raised across 1 of its 3 known rounds.