A long training session can span several shorter working contexts. That distinction matters when you plan an agent’s memory.
RadixArk released DaisyRose-27B on November 18, describing an agent-focused model built on Qwen3.8-27B. The release article reports reinforcement learning on about 1,200 curated tasks over 18 days using 144 H200 GPUs. The weights are available under Apache 2.0 in several formats.
The results make the release worth examining, particularly for teams studying how training tasks and infrastructure affect an agent. The documentation also contains a distinction that can disappear in a short headline: training sessions reached roughly one million tokens across trajectories, while the model’s native context window remains 256K tokens.
The longer session depends on how the surrounding agent handles its work and summarizes history. It does not mean a deployment can place a million tokens into one ordinary model context and expect them all to remain available.
Compare each reported gain with its own baseline
RadixArk’s table compares DaisyRose with the original Qwen3.8-27B model across several evaluations. The selected rows below add an absolute-change column, calculated from the published figures.
| Evaluation | Qwen3.8-27B reported score | DaisyRose-27B reported score | Absolute change |
|---|---|---|---|
| Terminal-Bench 4.1, pass@1 | 5.6% | 12.0% | +6.4 percentage points |
| Terminal Bench 2.1 | 73.0 | 75.5 | +2.5 points |
| DeepSWE v1.1 | 42.2 | 46.2 | +4.0 points |
| SWE Bench Pro v2 | 91.3 | 94.7 | +3.4 points |
These are developer-reported results. RadixArk says it selected the higher BF16 or FP8 score for each benchmark, with differences below one point wherever both formats were evaluated. That selection rule belongs beside the numbers when comparing them with another release.
The Terminal-Bench 4.1 change can also be expressed as a relative ratio: 12.0 divided by 5.6 is about 2.14. The absolute improvement is 6.4 percentage points. Calling it a 6.4% improvement would confuse the two quantities.
Neither expression establishes a success rate for a new repository. The benchmark version, task distribution, agent setup, and scoring rules remain part of the result. Scores from Terminal Bench 2.1 and Terminal-Bench 4.1 should not be treated as interchangeable measurements simply because the names are similar.
The task selection process is part of the story
RadixArk describes building candidates with Eigent AI and retaining about 1,200 tasks after review. It says the team examined base-model errors, adjusted task difficulty, and removed or corrected cases that rewarded the wrong behavior.
That account points to a useful question for teams planning their own training work: what precisely does the reward establish? A task can be easy to grade while measuring something different from the product’s goal. For example, a file’s existence may be easy to detect without establishing that its contents are useful.
Reviewing attempted solutions alongside the task can expose shortcuts. The task designer can then change the acceptance rule before using it as a training signal. RadixArk’s description provides a concrete example of that process, although the release alone does not establish that every outside team can reproduce the same gains.
Keep training evidence and release artifacts separate in any replication plan. Downloadable weights and serving guidance are inspectable outputs. Reproducing the training run also needs the relevant inputs, environment, task definitions, and evaluation settings. Check which parts are actually published before calling the full process reproducible.
A session, a trajectory, and a context window need separate labels
The release article explains that long training sessions can span multiple trajectories through harness summarization. The model card describes the model properties and inference settings, including the 256K native context window.
A context window is the material the model can use within the request’s supported limit. A longer agent session may instead use a sequence of requests, external files, summaries, and tool results. Its total token use can exceed what any individual request contains.
That design has a practical consequence. A summary preserves selected information while leaving out other details. If a later step needs an exact identifier or an earlier failed experiment, the agent needs a reliable way to recover it. A long session length does not, by itself, establish that such recovery works.
For a coding assistant, persistent artifacts can help. A concise task record can preserve the repository revision, decisions, checks, and open questions. The agent can reopen source files when it needs exact content rather than relying entirely on a compressed description of the past.
Read the training-resource figures as resource figures
RadixArk reports 18 days on 144 H200 GPUs and about 63,000 GPU-hours. Multiplying 18 by 24 by 144 gives 62,208 GPU-hours. The figures are consistent at the level of a rounded run description, but they do not supply a complete dollar cost.
A financial estimate would need the actual GPU pricing, scheduling, storage, networking, and other infrastructure costs. It would also need to establish what the reported hours include. Assigning a generic hourly price would produce a hypothetical budget, rather than the bill for this release.
The same distinction applies to the developer’s efficiency comparisons. Faster training or higher hardware utilization can help explain a system’s economics. Those measurements depend on the backend, hardware, workload, and accounting boundary being compared.
For a founder buying an inference service, training efficiency is background information. The immediate product decision concerns the price and quality of the served model on the application’s tasks. A cheaper training process does not automatically establish a cheaper endpoint.
Plan an evaluation that exposes lost state
A useful pilot can include tasks that require the agent to return to an earlier decision after several tool calls. Give each task an inspectable record of the original requirements and save the intermediate artifacts. Then compare the final output with that record.
Observe whether the agent preserves constraints across a summary, reopens the necessary source material, and correctly distinguishes a completed check from a planned one. These are proposed observations. We have not rerun DaisyRose’s training or benchmark harness.
Keep checkpoint format and runtime version in the result. The release offers BF16, FP8, and NVFP4 paths, and points to SGLang serving guidance. A result from one format and configuration should remain attached to that configuration.
DaisyRose is interesting because the public account connects task curation, training infrastructure, and reported agent gains. The useful next step is to preserve those distinctions when trying it: compare each benchmark on its own terms, treat 256K as the documented native context, and test whether the agent can recover the state its real work depends on.



