RadixArk’s DaisyRose-27B release asks how much targeted reinforcement learning can change a model’s performance on long agent tasks. The model card describes an RL-only adaptation of Qwen3.8-27B, with open weights and reported benchmark improvements. The useful comparison is the effect of the training process under a clearly specified evaluation setup, rather than the largest number in a cross-model table.
A targeted post-training account
The RadixArk model card describes approximately 1,200 curated long-horizon tasks, 18 days of training on 144 H200 GPUs and no supervised fine-tuning or distillation. It says the release preserves the base architecture, tokenizer and chat template. These are developer-reported properties of the release; this publication has not independently reproduced the training run.
The card reports gains on several coding, reasoning and instruction-following evaluations. It also documents important comparison qualifications: repeated runs, different harnesses in some external scores and the higher of BF16 and FP8 results for its own model. Some table entries use different benchmark versions or subsets. Those notes materially affect what a direct ranking can support.
Long tasks need more than a final score
An agent can fail through planning, tool use, context management or an incorrect final action. A total success rate does not show which part improved after training. Trace analysis on matched tasks would help identify whether DaisyRose makes better decisions, recovers more effectively or simply spends more inference on an attempt.
Training on a relatively small curated set can be effective when the examples exercise useful behaviors. The next question is transfer beyond the task distribution used during training. A held-out evaluation should include unfamiliar environments and variations that preserve the underlying problem without repeating the same surface form. That is a proposed verification method rather than an additional result in the release.
Inference conditions also belong in the comparison. Long context and high reasoning effort can affect both success and resource use. A model that completes more tasks at a larger budget may still be useful, but the operating tradeoff should be visible. Cost, elapsed time and retries should accompany task completion where a deployment decision depends on efficiency.
The deployment proposition also depends on compatibility in practice. Preserving the base interfaces can reduce migration work, but a changed policy may use tools or context differently. Existing applications should verify output handling and task behavior instead of assuming that architectural compatibility guarantees identical operating characteristics.
What a reproducible comparison would contain
A useful replication would pin the exact model revisions, serving precision, harness version, task subset and stopping rules. It would evaluate both models under the same conditions and report uncertainty across repeated attempts. External leaderboard numbers can provide context, but they should remain separate where the evaluation conditions cannot be aligned.
The review should inspect unsuccessful traces as well as successful ones. A higher aggregate score can coexist with a new failure mode that matters to a particular customer. Task-specific regression testing is therefore necessary before replacing a base model in an established workflow, especially where a long attempt can consume substantial resources.
The release establishes a concrete post-trained model and a detailed developer account of its evaluation. Its next persuasive milestone is a matched, independent comparison showing sustained gains on new long tasks at a stated operating budget. That would turn a promising training report into a clearer basis for deployment.
Sources and media credits
RadixArk’s DaisyRose-27B model artwork from the official weights repository. Model identity, not independent benchmark replication. Image: RadixArk


