RadixArk released DaisyRose-27B on November 18, building on Qwen3.8-27B with reinforcement learning. Weights are available under Apache 2.0. The author describes around 1,200 curated extended tasks and sessions of up to one million tokens using summarization. Native context remains 256K. This is an important boundary between the duration of a workflow and the size of one input.
What the published evaluation establishes
The model card reports Terminal-Bench 4.1 pass@1 of 12.0%, against 5.6% for the base model, using eight repetitions per task and an internal execution environment. This is the author's evaluation, not our comparative test. The 6.4 percentage-point difference is not universal reliability across arbitrary tasks.
For someone assigning project work to an agent, the whole process matters: preserved context, available tools, verification and a clear completion point. A model can perform an individual step well and still lose an important condition later. This is a general engineering consideration, not a failure we have observed in DaisyRose.
Memory requirements extend beyond weights
The card gives approximate weight sizes of 52 GB for BF16, 29 GB for FP8 and 23 GB for NVFP4. A running service has additional requirements. Context, cache and concurrent requests must fit alongside the model; file size is not a complete hardware guarantee.
Our pilot proposal is to select a real, authorized task with a known success criterion. For a code change, that might include a reviewed diff, applicable tests and no unrequested modifications. For document analysis, it could mean preserving essential conditions and explicitly identifying contradictions.
Count unsuccessful attempts too
A fair comparison gives the base model and new candidate equal budgets and tools. Record terminated tasks as well as the strongest demonstration. Unlimited retries can hide a cost the real workflow cannot sustain.
Extended tasks also deserve a review of intermediate summaries. It is useful to distinguish a wrong decision from lost context or a failed tool. That makes evaluation actionable for the complete system rather than merely a ranking of models.
Released weights enable these local comparisons. SIExpo has not loaded DaisyRose and does not promise performance on a particular computer. The verified news is the open model and the described training and evaluation approach.



