What changed
RadixArk released Apache-2.0 BF16, FP8 and NVFP4 checkpoints for DaisyRose-27B, a reinforcement-learning post-train of Qwen3.8-27B. The model retains image input and a 256K-token context window, and the release includes SGLang serving guidance for its long-horizon agent configuration.
Why it matters
A runnable 27B checkpoint turns the post-training claim into something teams can test on private agent workloads; the published weights and serving recipe establish availability, not the claimed benchmark or training-efficiency gains.
What to do
Pin one checkpoint and the documented SGLang configuration, then compare it with Qwen3.8-27B on the same long-horizon tasks, harness, tool budget and hardware while logging failures, latency, token use and cost.
Caveat
RadixArk authored the benchmark, data-selection and training-efficiency evidence. The reported gains are not independently reproduced, scores use release-specific harnesses and repeated-run rules, and the model inherits limitations of its base model and training data.