LIVE

RadixArk Unveils DaisyRose-27B, Focused on Agentic Efficiency

🚀 Startups & Funding 2 min read 👁 0

Startup RadixArk has launched DaisyRose-27B, an AI model optimized for long-horizon agent tasks, claiming significant improvements in training efficiency and reasoning capabilities.

RadixArk Unveils DaisyRose-27B, Focused on Agentic Efficiency🚀Startups & Funding

Photo: huggingface.co

Key points

  • DaisyRose-27B utilizes a 256K-token context window and is optimized for long-horizon agent tasks.
  • RadixArk claims a 57% reduction in training costs compared to traditional open-source training backends.
  • The model showed a performance jump on Terminal-Bench 4.1, doubling its pass@1 rate.

RadixArk has introduced its inaugural AI model, DaisyRose-27B, which is built upon the Qwen3.8-27B architecture. The model was trained using a reinforcement learning framework dubbed Miles, specifically targeting complex agent-based tasks that require processing sessions up to 1 million tokens. According to the company, the model demonstrated a notable performance increase on Terminal-Bench 4.1, with pass@1 rates rising from 5.6% to 12.0%.

The training process utilized 144 H200 GPUs over an 18-day period, relying on a curated set of 1,200 tasks developed in collaboration with Eigent AI. RadixArk reports that its proprietary training backend achieved 37% model FLOPs utilization, which it claims is significantly more efficient than standard open-source alternatives. This architecture allowed for a 57% reduction in GPU training costs and maintained stability over 63,000 total GPU-hours.

Technical Highlights:

  • Maintains the 256K-token context window and image understanding capabilities of the base Qwen model.
  • Employs an asynchronous training loop where the agent and trainer operate simultaneously without waiting for one another.
  • Utilizes the AgentENV project for sandbox environments, which the company claims reduced cloud infrastructure costs by 80% compared to typical providers.

The model is designed to think before providing an answer, separating reasoning content from the final output. RadixArk recommends deploying the model using SGLang and DSpark speculative decoding for optimal performance.

Source huggingface.co — huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post… ↗

Contact the newsroom

Got a tip, a proposal or a partnership question? Write to us and we will reply shortly.