RadixArk releases DaisyRose-27B: 1,200 tasks and 18 days of training lift Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%
Synopsis
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks using only about 1,200 curated tasks over 18 days on 144 H200 GPUs. On Terminal-Bench 4.1 its pass@1 rises from the base model's 5.6% to 12.0%, and it also scores above the base on every other listed reasoning, instruction-following and coding benchmark. The model keeps the base architecture, tokenizer, 256K context window and image understanding, with checkpoints released under Apache 2.0.
Interpretation
A 27B-scale model gains a doubling of pass@1 on long-horizon terminal tasks from about 1,200 curated tasks and 18 days of reinforcement-learning post-training. The base Qwen3.8-27B scores 5.6% pass@1 on the same benchmark and DaisyRose-27B scores 12.0%, the largest gain in the comparison. In the same benchmark table DaisyRose-27B scores 75.5 versus 73.0 on Terminal Bench 2.1, 46.2 versus 42.2 on DeepSWE v1.1, 94.7 versus 91.3 on SWE Bench Pro v2, 34.3 versus 33.9 on text-only HLE, 82.4 versus 79.5 on IFBench and 90.7 versus 89.2 on GPQA Diamond, all above the base.
Training tasks were redesigned around mistakes the base model makes, with difficulty adjusted until the solve rate fell in a useful learning range. Tasks came from thousands of candidates built with Eigent AI from its large pool in machine learning, software engineering, science and other fields, and about 1,200 entered the final training set. Each task requires several interdependent steps, with each attempt getting up to 1.5 hours in a sandbox, more than 100 tool calls and more than 256K output tokens; every task was reviewed alongside the model's attempts, and tasks that could be passed without solving the intended problem or that had design flaws were fixed or removed.
The training backend runs fully asynchronously, so the agent keeps working on tasks while the trainer keeps learning from trajectories and neither waits for the other. 128 H200s run inference on agent tasks while 16 more run the trainer, and after every step the updated model goes back out to those 128 GPUs. On this backend 37% of each GPU's peak compute went into useful training work (model FLOPs utilization, or MFU), against 16% for the open-source Megatron-based trainer on comparable runs, so the same training needs 57% less GPU time; end to end the release run trained at about 16,800 tokens per second versus about 8,300 for the open-source trainer on the same GPUs.
Perspective
The report covers RadixArk's DaisyRose-27B and its training configuration: the base is Qwen3.8-27B, training ran on 144 H200s for 63,000 GPU-hours in total, and checkpoints are offered under Apache 2.0 in BF16, FP8 and NVFP4, with SGLang plus DSpark speculative decoding recommended for serving.
Benchmark scores use the higher of the BF16 and FP8 runs, and the authors state the two differ by less than one point wherever both were evaluated; the open-source trainer figures in the efficiency comparison come from comparable runs rather than the release run itself; sandbox creation succeeded 99% of the time across nine weeks and above 99.95% since late October.
