RadixArk Cuts AI Post-Training GPU Costs by 57% With DaisyRose-27B
RadixArk used an asynchronous reinforcement learning setup across 144 H200 GPUs to train its new DaisyRose-27B model, more than doubling benchmark task performance.
RadixArk has released DaisyRose-27B, its first artificial intelligence model. The system is built on top of Qwen3.8-27B and uses reinforcement learningReinforcement learningA machine learning method where software learns by trial and error using rewards and penalties. Open in glossary to improve how agents handle complex, multi-step computer tasks. Over an 18-day trainingTrainingTeaching an AI model by showing it very large amounts of data. Open in glossary run using 144 Nvidia H200 GPUs, the model demonstrated major performance jumps across several developer benchmarks.
Targeted Training for Long Sessions
The model was trained on 1,200 curated tasks created in partnership with Eigent AI. These challenges spanned machine learning, software engineering, and scientific problems. RadixArk exposed the model to extended task sessions containing up to one million tokens, using a method called harness summarization to manage the context across several attempts.
DaisyRose-27B keeps the base architecture of Qwen3.8-27B, including its tokenizer, image understanding, and 256,000-token context windowContext windowHow much text an AI model can keep in view at one time. Open in glossary. By default, the model works through reasoning steps before delivering its final answer. On the Terminal-Bench 4.1 test, which measures how well models perform in command-line environments, DaisyRose-27B lifted the pass rate from 5.6% to 12.0%.
Hardware Efficiency and Sandboxing
The post-training run used RadixArk’s open-source framework, Miles, operating asynchronously. In this setup, 128 GPUs generated task runs while 16 GPUs continuously updated the model parameters without waiting for each other. This workflow achieved a model FLOPs utilization of 37%, compared to 16% for a standard setup using Megatron. The team claims this architecture cut needed GPUGPUA chip that does many small sums at once. First made for graphics, now the main chip for AI. Open in glossary training time by 57% and processed roughly 16,800 tokens per second.
To let the agent test code safely, RadixArk managed task environments with an open-source tool called AgentENV. The training run generated around 944,000 separate sandboxSandboxAn isolated digital space used to run and test untrusted software safely. Open in glossary environments over nine weeks, peaking at 1,566 running at the same time. The team lowered infrastructure costs by about 80% compared to standard cloud providers by dynamically scaling memory for each sandbox across up to 12 host machines.
How does this story land?
One tap per reader.
Quick poll
Will asynchronous agent training become standard for post-training open models?
One tap. No sign-up.
Ask this story
Still have a question?
Answers use only what is in this story. They are written by AI and can be wrong.