RadixArk released Miles v0.1 to align AI training and inference
RadixArk's release builds on Miles, first introduced in November 2025, while its $100M seed funds an open-source AI infrastructure strategy.
By RuntimeWire Staff · Published
Primary source: MiniMax
Why it matters
RadixArk is trying to turn SGLang and Miles into a managed infrastructure business after raising $100M at a $400M post-money valuation. Miles v0.1 gives founders Ying Sheng and Banghua Zhu another release around which to build that commercial layer.
Ying Sheng (@ying11231) and Banghua Zhu (@BanghuaZ) released Miles v0.1 through RadixArk, the company they founded to commercialize open-source infrastructure for AI inference and post-training.
Miles is one of the two open-source foundations Sheng and Zhu are turning into a commercial infrastructure business. SGLang handles model inference and rollouts; Miles handles reinforcement learning and post-training. RadixArk is building managed products around both projects.
Sheng, RadixArk's co-founder and CEO, contributed to SGLang. Zhu, RadixArk's co-founder and CTO, earned his PhD in electrical engineering and computer sciences at UC Berkeley, previously co-founded Nexusflow AI and is an incoming University of Washington faculty member. Zhu describes RadixArk's mission as bringing frontier-level AI infrastructure to a wider group of builders through systems developed in the open.
A version number for an existing project
The v0.1 label should not be read as Miles' first public appearance. LMSYS introduced Miles on November 19, 2025, and RadixArk presented Miles as a core project when RadixArk formally launched on May 5, 2026. Its documentation covers low-precision training, speculative rollouts, model support and distributed reinforcement learning.
In the latest supplied repository crawl, the Miles repository had approximately 1,400 GitHub stars. Its documentation describes Miles as an enterprise-oriented framework that pairs SGLang's inference engine with NVIDIA's Megatron-LM training system. A PyTorch FSDP2 backend is also available, while RadixArk directs its largest-model recipes toward Megatron-LM.
Miles is designed around a recurring systems problem in reinforcement learning: training and inference can behave differently even when they are nominally running the same model. Changes in numerical precision, mixture-of-experts routing and runtime implementation can distort the probabilities used to update a model. Miles includes low-precision training workflows, routing replay and mechanisms for moving updated weights back into inference engines during a running training loop.
That puts Miles in the same working category as open-source post-training projects including slime, from which Miles was forked, as well as OpenRLHF and verl. RadixArk's stated distinction is the tight coupling between SGLang rollouts and large-scale training. The architecture is intended for research groups and model labs running expensive, distributed workloads where small runtime inconsistencies can derail an entire training run.
The $100M open-source business model
Sheng and Zhu formally launched RadixArk on May 5 with a $100 million seed round at a $400 million post-money valuation. Accel led the financing and Spark Capital co-led it. NVentures, Salience Capital, A&E Investment, HOF Capital, Walden Catalyst, AMD, LDV Partners, WTT Fubon Family, MediaTek and Databricks also participated.
RadixArk said the capital would support SGLang and Miles, add model and hardware coverage, and fund managed infrastructure for organizations operating AI systems at scale. Open-source adoption supplies distribution for the managed products RadixArk plans to sell.
The founders have picked a familiar infrastructure strategy with unusually large seed financing behind it. RadixArk can give away the core software, build trust through public engineering work and charge customers that need deployment help, guaranteed performance or managed capacity. The hard part is turning open-source repositories into durable revenue without weakening the projects that created the distribution.
Miles v0.1 is an incremental marker in a project introduced publicly nine months earlier. Its importance comes from RadixArk taking shape around it. Sheng and Zhu are betting that maintainers of critical open-source AI systems can own the commercial layer above those systems. The release gives RadixArk another defined milestone as it tries to convert that engineering work into a managed infrastructure business.