Artificial Intelligence1 min read

DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

By · Published by Everything Blog

In short

DaisyRose-27B was developed through an 18-day reinforcement learning run on 144 H200 GPUs using Qwen3.8-27B as its foundation. The training leveraged 1,200 curated tasks created with Eigent AI across science, machine learning, and software engineering, requiring models to navigate chained dependencies and up to 1.5 hours of sandbox runtime. RadixArk supported the training effort using an internal sandbox system built on AgentENV that executed roughly 944,000 sandboxes between mid-September and mid-November.

Key points

  • The internal training backend reached 37% model FLOPs utilization, cutting the necessary…: The internal training backend reached 37% model FLOPs utilization, cutting the necessary GPU time by 57% compared to Megatron-based alternatives.
  • The 63,000 GPU-hour run processed roughly 16,800 tokens per second, delivering an end-to-…: The 63,000 GPU-hour run processed roughly 16,800 tokens per second, delivering an end-to-end speed twice as fast as open-source trainers.
  • Public checkpoint releases in BF16, FP8, and NVFP4 formats expand access to models tuned…: Public checkpoint releases in BF16, FP8, and NVFP4 formats expand access to models tuned for long-horizon agent execution.

Sources checked by our analysis pipeline: 1

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

Today we are releasing DaisyRose-27B, RadixArk’s first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles. We trained it on long-horizon agent tasks with sessions up to 1M tokens, with just 1,200 curated tasks in 18 days on 144 H200 GPUs. DaisyRose-27B improves Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. Beyond Terminal-Bench 4.1, it also delivers gains on a variety of other reasoning, instruction following, and coding benchmarks.

We are releasing BF16, FP8, and NVFP4 checkpoints for everyone to try and build on. In this post, we explain how carefully designed tasks and efficient infrastructure made these gains possible on a limited compute budget.

DaisyRose-27B scores higher than the original Qwen3.8-27B model on every benchmark above1. The largest gain is on Terminal Bench 4.1, a notoriously hard long-horizon benchmark. Our magic lies in the following three key ingredients: high-quality training data, efficient training backend, and robust sandboxes.

High-quality Training Data

We built the tasks in partnership with Eigent AI, starting from Eigent AI’s large pool of tasks in machine learning, software…

Artificial Intelligence