RadixArk has released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on long-horizon agent tasks. The headline result is a jump on Terminal-Bench 4.1, where pass@1 more than doubles from 5.6% to 12.0%.

The model was trained on just 1,200 curated tasks over 18 days on 144 H200 GPUs, with agent sessions running up to 1 million tokens. It keeps the base model's architecture, tokenizer, chat template, image understanding and 256K-token context window, and it beats Qwen3.8-27B on every benchmark RadixArk reports, including Terminal-Bench 2.1 (73.0 to 75.5), DeepSWE v1.1 (42.2 to 46.2), SWE-Bench Pro v2 (91.3 to 94.7) and GPQA Diamond (89.2 to 90.7).

Training used Miles, RadixArk's open-source reinforcement learning framework, running in fully asynchronous mode: 128 H200 GPUs ran inference on agent tasks while 16 more trained on the resulting trajectories. RadixArk says the setup reached 37% model FLOPs utilization, against 16% for the open-source trainer it compared against, cutting GPU time for the same training by 57%.

The checkpoints are available now under Apache 2.0 in BF16, FP8 and NVFP4 formats, and RadixArk recommends serving the model with SGLang and DSpark speculative decoding.