Ex-xAI & NVIDIA Engineers Build DaisyRose-27B on 144 H200s with Custom RL Infra
RadixArk, founded by ex-xAI inference lead Sheng Ying and ex-NVIDIA researcher Zhu Banghua, releases DaisyRose-27B, a 27B parameter model trained on 144 H200 GPUs in 18 days using their custom RL framework Miles and SGLang inference engine, achieving 2.14x pass@1 improvement on Terminal-Bench 4.1.
Founders' Background and Motivation
RadixArk was co-founded by CEO Sheng Ying and CTO Zhu Banghua, both born in the 1990s. Sheng Ying is a co-creator of SGLang, an open-source inference engine now running on hundreds of thousands of GPUs and used in production at Meta, Microsoft, and Google. She earned a PhD from Stanford under Clark Barrett, winning best-paper awards at IJCAR 2020 and TACAS 2022 for formal verification work, then shifted to LLM systems, contributing to FlexGen, S-LoRA, FastChat, and Vicuna. After a brief stint co-leading the inference team at xAI (Oct 2024 – Aug 2025), she left to focus full-time on SGLang's growing community needs and to preserve the open-source culture she experienced at Databricks.
Zhu Banghua graduated top of his class (1st out of 262) in Electronic Engineering at Tsinghua University, then pursued a PhD at UC Berkeley under Jianqiao Zhu and Michael I. Jordan, focusing on efficient LLM fine-tuning and serving. He won the 2023 David J. Sakrison Memorial Prize for outstanding doctoral research. His open-source contributions include Starling-7B, Athene-V2-72B, and the evaluation platforms Chatbot Arena and Arena-Hard-Auto. He co-founded Nexusflow (enterprise AI agents) in 2023, which was acquired by NVIDIA in 2025; he joined as a Principal Research Scientist on the Star Nemotron team but left within six months, forfeiting unvested acquisition proceeds, to start RadixArk with Sheng Ying.
The two had previously collaborated on LLM inference caching and the S-LoRA project. RadixArk now centers development on the SGLang inference engine and the Miles reinforcement-learning framework, the latter used to train their first model.
DaisyRose-27B: Training Setup and Infrastructure
Released on November 18, DaisyRose-27B is based on Qwen2.5-27B (referred to as Qwen3.8-27B in the article). Training ran for 18 days on 144 NVIDIA H200 GPUs using the Miles RL framework. The team's primary goal was to verify that million-token long-horizon agent training could run stably while improving GPU utilization and reducing sandbox costs. They curated ~1,200 long-horizon agent tasks with Eigent AI, adjusting difficulty to the base model's failure modes and manually auditing execution traces to fix or remove tasks with reward loopholes.
Each task attempt ran up to 1.5 hours, involved over 100 tool calls, and some sessions accumulated 1 million tokens. Trajectories were organized via a Harness summarization mechanism. Training used Miles' fully asynchronous mode: 128 H200s executed tasks and generated rollouts while 16 H200s performed model updates, enabling parallel environment interaction and gradient steps.
Performance Metrics and Efficiency Gains
Internal measurements report a model compute utilization (MFU) of 37% for the training backend, versus 16% for a comparable Megatron-based open-source trainer. This translates to roughly 57% less GPU time to complete the same training. End-to-end training throughput (including agent execution time) reached ~16,800 tokens/second, about 2× the open-source baseline. Total compute consumed was approximately 63,000 GPU-hours, with stable speed and utilization throughout the 18-day run.
On Terminal-Bench 4.1, DaisyRose-27B's pass@1 rose from the base model's 5.6% to 12.0%, a 2.14× improvement. The model targets multi-step reasoning, planning, and tool-use for long-horizon agents.
Deployment and Sandbox Infrastructure
RadixArk released model weights in BF16, FP8, and NVFP4 precisions. SGLang added same-day support for deployment. For large-scale task execution, the team built a sandbox service on the open-source AgentENV project. From mid-September to mid-November it ran ~944,000 sandboxes, peaking at 1,566 concurrent instances. Since late October, sandbox creation success rate exceeded 99.95%.
Future Direction
RadixArk plans to offer post-training support to teams building specialized models, aiming to enhance target capabilities while preserving general competence. DaisyRose-27B serves as a demonstration of their integrated stack — inference, RL, and sandbox infrastructure — to help other teams train and run models.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
