RadixArk has released DaisyRose-27B, its first openly downloadable model, aiming at a persistent weakness of coding agents: finishing long, multi-step jobs rather than merely writing plausible snippets. Built on the existing Qwen3.8-27B architecture, the checkpoint has been post-trained with reinforcement learning on agent tasks. The developer provides BF16, FP8 and NVFP4 variants and declares Apache-2.0 licensing. These are new derivative weights, not a freshly pretrained 27-billion-parameter foundation model. Primary source ↗
The technical story is the unusually deliberate training curriculum. RadixArk and Eigent AI started with thousands of candidate problems in software engineering, machine learning and science, then selected about 1,200 tasks where the base model made identifiable mistakes. Each exercise was designed to require multiple dependent actions and to reward completion of the real objective. RadixArk says attempts could use sandboxes for up to 90 minutes and more than 100 tool calls. This is more specific than a generic instruction-tuning dataset.
Training lasted 18 days on 144 H200 GPUs, with 128 GPUs running agent inference and 16 used for learning, according to the release. Some training sessions spanned up to one million tokens through summarization across trajectories. That must not be confused with an inference context of one million tokens: DaisyRose keeps Qwen3.8-27B's advertised 256K-token context window and multimodal input pathway. The model card recommends SGLang and optional speculative decoding components for serving.
The developer's own evaluation puts Terminal-Bench 4.1 pass@1 at 12.0% versus 5.6% for the original Qwen model, and lists gains on several other coding and reasoning tests. The published comparison relies on specified harnesses, repetition counts and grading choices. These values are useful primary-source evidence of the team's claims, not independent Open Model Weights measurements or proof that DaisyRose is universally superior to competitors.
RadixArk also says its internal training backend reduced GPU time relative to a selected open trainer configuration by 57%, while supporting long-lived sandbox jobs and asynchronous learning. That number is workload- and baseline-specific, not a general statement that any team can cut its reinforcement-learning bill by half. Training infrastructure, data quality and reward design are part of the result.
For developers, this release offers a concrete experiment: pin a particular checkpoint, check its license and safetensors manifest, then reproduce agent evaluations on the actual harness and tasks that matter to the deployment. The larger lesson is that focused, verifiable post-training may change agent behavior substantially without changing the base model's size or claiming a new architecture.
What this changes for the evidence layer
New candidate, not a source-pinned catalog record. Inspect official BF16, FP8 and NVFP4 file manifests and revision SHA independently; check license file, model lineage and required serving components. Benchmarks and training-efficiency claims are RadixArk's results, not OMW measurements; do not grant Leading curation from reported scores alone.
RadixArk — official DaisyRose-27B release report
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗