RadixArk has released DaisyRose-27B, its first model, a post-trained version of Qwen3.8-27B. The company trained it with reinforcement learning on long-horizon agent tasks, with sessions up to 1 million tokens, using 1,200 curated tasks, its open-source Miles framework and 144 H200 GPUs. The checkpoints are available under the Apache 2.0 licence in BF16, FP8 and NVFP4 formats. The model keeps the base model's architecture, tokenizer, image understanding and 256K-token context window, and the company recommends serving it with SGLang and DSpark speculative decoding. The headline result is on Terminal-Bench 4.1, where pass@1 rose from 5.6% to 12.0%. RadixArk also reports gains over the original model on 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), IFBench (79.5 to 82.4), GPQA Diamond (89.2 to 90.7) and HLE text-only (33.9 to 34.3). The scores are company-reported, use the higher of the BF16 and FP8 runs, and no independent evaluation is cited. RadixArk also says its training backend reached 37% model FLOPs utilisation against 16% for the open-source Megatron-based trainer on comparable runs. That cut GPU time by 57%, and end-to-end throughput roughly doubled, to about 16,800 tokens per second against 8,300. The release run used 63,000 GPU-hours over several weeks. For Australian businesses and professionals, the immediate impact is limited. The weights can be downloaded free under Apache 2.0, but running a 27B model at this throughput needs data-centre-class GPUs, and the training efficiency gains matter mainly to teams that train their own models. Businesses using hosted AI services see no direct change from this release. Any organisation that sends personal information to a model, whether self-hosted or not, still has to meet its obligations under the Privacy Act 1988, including the Australian Privacy Principles on security and cross-border disclosure. It is worth testing only if you build coding or IT automation agents, and the benchmark gains should be checked against your own tasks first.