RadixArk

RadixArk is an American artificial intelligence infrastructure startup founded in 2025 by Ying Sheng and Banghua Zhu, the creators and core maintainers of SGLang, the open-source large-language-model inference engine. The company commercializes SGLang along with Miles, an open-source reinforcement learning framework, providing managed infrastructure and tooling for inference and training workloads. As of May 2026, RadixArk has raised $100 million in seed funding at a $400 million post-money valuation, in a round led by Accel and co-led by Spark Capital.

At a glance

• Founded: 2025 in the United States. Public launch and seed-round announcement in May 2026.
• Status: Private. Seed round closed May 2026.
• Funding: $100 million seed at $400 million post-money valuation. Led by Accel; co-led by Spark Capital. NVentures (NVIDIA's venture capital arm), Salience Capital, A&E Investments, HOF Capital, Walden Catalyst Ventures, AMD, LDV Partners, WTT Investment, and MediaTek participated. Angel investors include Igor Babuschkin (co-founder of xAI), Lip-Bu Tan (CEO of Intel), Hock Tan (CEO of Broadcom), John Schulman (co-founder of OpenAI and Thinking Machines Lab), Soumith Chintala (creator of PyTorch), and Thomas Wolf (co-founder of Hugging Face).
• CEO: Ying Sheng, co-founder. Co-creator of SGLang and previously a researcher at xAI. Earlier contributions to LMSYS Chatbot Arena and to FlexGen.
• Other notable leadership: Banghua Zhu, co-founder. Co-creator of SGLang and previously a researcher at NVIDIA. Long-term collaborator with Sheng dating to their PhD work.
• Open weights: Not applicable. RadixArk develops infrastructure for inference and reinforcement learning rather than foundation models. SGLang and Miles are Apache 2.0 licensed.
• Flagship products: SGLang (open-source inference engine), Miles (open-source reinforcement learning framework), and RadixArk's commercial managed-infrastructure offerings on top of those open cores.

Origins

Ying Sheng and Banghua Zhu collaborated as PhD researchers in the AI infrastructure space and co-led a series of high-impact open-source projects. SGLang originated at LMSYS, the research collective at Berkeley known for the Chatbot Arena leaderboard (LMArena) and for releases including Vicuna and the LMSYS Chatbot Arena. The SGLang project began as a serving-framework research effort in 2023 and 2024, with subsequent contributions from a broad community of researchers and practitioners across academic and industry labs.

The technical core of SGLang is RadixAttention, a method for caching and reusing the prefix portions of language model computation across requests, combined with a structured-program-execution model that compiles structured generation patterns (regex-constrained outputs, structured JSON, agentic control flow) into efficient runtime sequences. The combination produces high-throughput serving for the structured-generation workloads that have become characteristic of agentic AI applications.

By 2025, SGLang had become one of the de facto standard inference engines for production-scale LLM deployment, alongside vLLM. The project is reported to power inference at Google, Microsoft, NVIDIA, Oracle, AMD, Nebius, LinkedIn, xAI, Thinking Machines Lab, and additional operators, with trillions of tokens per day flowing through SGLang-powered deployments. xAI's Igor Babuschkin has publicly stated that SGLang is xAI's default inference engine.

The RadixArk spinout was announced on May 5, 2026, with the $100 million seed round led by Accel and co-led by Spark Capital. The company commercializes SGLang along with Miles, a separately-developed open-source reinforcement learning framework. The "open-core plus managed infrastructure" framing is structurally similar to the Inferact approach to vLLM, with both companies launching within four months of each other.

Mission and strategy

RadixArk's stated mission is to "grow SGLang and democratize frontier AI infrastructure." The strategic premise reflects two threads. The first is that SGLang and Miles, as widely-adopted open-source infrastructure components, benefit from commercial backing and that managed infrastructure on top of these open cores is a viable enterprise commercial model. The second is that RL and inference are converging into a unified workflow ("post-training and inference") and that a vendor with strength across both layers is positioned to capture the larger value than either alone.

The strategy combines three threads. The first is continued investment in SGLang and Miles as community-governed open-source projects, with RadixArk's senior engineering staff providing principal maintainer attention. The second is the development of managed inference and training services for enterprise and frontier-AI-lab customers. The third is cross-vendor hardware optimization, with the participation of NVIDIA's NVentures, AMD as a strategic investor, and MediaTek as another semiconductor partner reflecting the cross-vendor neutrality of the open-source positioning.

The competitive premise is that the AI inference market supports multiple open-core platforms (vLLM and SGLang in particular), each with distinctive technical strengths and adoption patterns. RadixArk's positioning emphasizes structured-generation and agentic workloads where SGLang's RadixAttention and structured-execution capabilities produce the largest gains.

Models and products

• SGLang (open-source). High-performance serving framework for large language models and multimodal models. Apache 2.0 licensed.
• Miles (open-source). Reinforcement learning framework, also Apache 2.0 licensed. Targets the RL training workloads that have become central to post-training pipelines for reasoning and agentic models.
• RadixArk managed services. Commercial managed-inference and managed-training offerings on top of SGLang and Miles. Specific product details and pricing have not been broadly publicly disclosed as of May 2026.
• Cross-hardware optimization. Continued SGLang optimization for NVIDIA, AMD, and other AI hardware platforms.

Distribution is split between the open-source community (via GitHub for SGLang and Miles) and direct enterprise sales for the RadixArk commercial managed-service offerings.

Benchmarks and standing

SGLang is consistently ranked among the leading open-source LLM inference engines on community benchmarks for throughput, structured-generation efficiency, and prefix-caching performance. The project's distinctive technical advantage is on workloads with structured outputs (JSON-constrained generation, regex-constrained generation, agentic control flow), where RadixAttention produces the largest gains over alternative engines.

The trillions-of-tokens-per-day deployment count and the use of SGLang as the default engine at xAI and across major frontier-lab deployments are the clearest indicators of standing. Public adoption among Google, Microsoft, NVIDIA, Oracle, AMD, Nebius, LinkedIn, Thinking Machines Lab, and additional operators reflects the project's broad cross-organization usage.

For RadixArk specifically, commercial-traction data has not been publicly disclosed. The standing of the commercial entity will depend on the conversion of the open-source community position into managed-service revenue.

Leadership

As of May 2026, RadixArk's senior leadership includes:

• Ying Sheng, Chief Executive Officer and co-founder. Co-creator and lead architect of SGLang. Previously a researcher at xAI; earlier contributions to FlexGen and to LMSYS Chatbot Arena.
• Banghua Zhu, co-founder. Co-creator of SGLang. Previously a researcher at NVIDIA. Long-term research collaborator with Sheng since their PhD work.

The company's senior research-and-engineering bench draws on the broader SGLang and LMSYS communities, including contributors who have worked on Chatbot Arena and other LMSYS infrastructure. Specific senior-leadership titles beyond the co-founders have not been publicly disclosed.

Funding and backers

RadixArk's funding history through May 2026 consists of one disclosed round:

• Seed (May 2026): $100 million at $400 million post-money valuation. Led by Accel; co-led by Spark Capital. NVentures (NVIDIA), Salience Capital, A&E Investments, HOF Capital, Walden Catalyst Ventures, AMD, LDV Partners, WTT Investment, and MediaTek participated. Angel investors include Igor Babuschkin (xAI), Lip-Bu Tan (Intel), Hock Tan (Broadcom), John Schulman (OpenAI and Thinking Machines Lab), Soumith Chintala (PyTorch and Thinking Machines Lab), and Thomas Wolf (Hugging Face).

The seed-round configuration is unusual for the breadth of strategic-investor participation. The simultaneous involvement of NVIDIA (NVentures), AMD, MediaTek, Intel (via Lip-Bu Tan), and Broadcom (via Hock Tan) reflects the cross-vendor positioning of SGLang as hardware-neutral inference infrastructure. The angel-investor list, including senior frontier-AI figures from xAI, OpenAI, Thinking Machines Lab, and Hugging Face, reinforces the open-source community-anchored positioning.

The investor base spans US growth-stage VCs (Accel, Spark Capital), corporate-strategic investors (NVIDIA, AMD, MediaTek), and frontier-AI angel investors. The configuration is structurally distinct from peer inference startup Inferact, which closed its seed round at a higher valuation ($800 million) with a more concentrated VC-led investor base.

Industry position

RadixArk occupies a structurally distinctive position in the AI inference category as the principal commercial entity backing SGLang. The combination of the SGLang open-source project's adoption depth at frontier labs, the founder team's research-and-maintainer status, and the unusually broad strategic-and-angel investor base produces a profile that no peer commercial inference startup matches at the same combination of attributes.

Industry coverage has characterized RadixArk as one of two principal open-core commercialization plays in the AI inference category, alongside Inferact. Both companies launched within four months of each other; both target the open-core managed-service commercial model; and the principal differentiation between them is the underlying open-source project and the configuration of the strategic-investor base.

The principal strategic-execution risks identified are the conversion of the SGLang open-source community into managed-service customers, the maintenance of cross-vendor neutrality given the strategic-investor base, and the competitive pressure from peer inference platforms.

Competitive landscape

RadixArk competes for AI inference workload across several categories:

• Inferact. The closest direct peer, commercializing vLLM with a $150 million seed at $800 million post-money valuation in January 2026.
• Together AI, Fireworks AI. Direct AI inference platform competitors with their own optimized engines and managed services.
• Cloud-provider managed inference. Amazon Bedrock, Google Vertex AI, and Azure OpenAI Service are the principal cloud-provider alternatives.
• Cerebras, Groq, SambaNova. Specialized AI hardware vendors with their own inference stacks.
• NVIDIA inference software. TensorRT-LLM and Triton Inference Server are the principal NVIDIA-aligned alternatives. NVIDIA's NVentures participation in RadixArk's seed reflects the strategic-partnership rather than direct-competition framing.

Outlook

Several open questions affect RadixArk's trajectory in 2026 and 2027:

• Commercial managed-service traction, with named enterprise customers and revenue trajectory among the watchable signals.
• The competitive dynamic with Inferact and the broader question of whether the inference category supports multiple open-core commercial entities at scale.
• Continued SGLang technical leadership, particularly on agentic and reasoning workloads where structured-generation efficiency matters most.
• The development of the Miles RL framework as a parallel commercial vector.
• Subsequent fundraising; the $400 million seed valuation implies a meaningful step-up in commercial milestones before the next priced round.
• The cross-vendor neutrality of the company's hardware positioning, given the simultaneous NVIDIA, AMD, MediaTek, Intel, and Broadcom strategic-investor engagements.

Sources

• BusinessWire: RadixArk Launches with $100 Million in Seed Funding Led by Accel. Launch announcement.
• TechCrunch: Project SGLang spins out as RadixArk with $400M valuation. Pre-launch reporting.
• Accel: Investing in RadixArk. Lead-investor announcement.
• GitHub: sgl-project/sglang. Open-source project repository.
• HOF Capital: Why We Invested in RadixArk. Co-investor perspective.
• Tech Funding News: NVIDIA and Accel pour $100M into RadixArk. Strategic-investor context.