RadixArk's DaisyRose-27B Doubles Terminal Task Success With Reinforcement Learning
RadixArk's first open-weight model takes Qwen3.8-27B and more than doubles its Terminal-Bench 4.1 score using pure RL on 1,200 agent tasks.
Takeaways
- RadixArk released DaisyRose-27B, a Qwen3.8-27B RL post-train under Apache-2.0.
- Terminal-Bench 4.1 pass@1 jumped from 5.6% to 12.0% after RL.
- Trained 18 days on 144 H200s, roughly 63,000 GPU-hours.
- Uses 1,200 curated long-horizon agent tasks, sessions up to 1M tokens.
- RadixArk's Miles backend reports 37% MFU vs 16% for an open-source trainer.
- Third-party data-analytics test showed a 4-point regression against the Qwen base.
RadixArk has released its first open-weight model, DaisyRose-27B, a Qwen3.8-27B checkpoint post-trained exclusively with reinforcement learning on long-horizon agent tasks. The company published the model, quantizations, and training framework under Apache-2.0.
The checkpoint also demonstrates RadixArk’s training stack. DaisyRose was trained with Miles, its open-source reinforcement learning software, plus an internal backend that reportedly converted 37% of theoretical accelerator capacity into model computation. RadixArk measured 16% for a popular open-source trainer under the same test conditions.
A Qwen checkpoint tuned for agents
DaisyRose retains the base model’s tokenizer, chat template, vision encoder, and 262,144-token context window. Existing Qwen3.8 inference and post-training pipelines should generally load it without template changes.
Post-training used reinforcement learning, which updates behavior from reward signals, without supervised fine-tuning on curated answer pairs or knowledge distillation from a teacher model. The release has the following specifications:
- Parameters: 27.8 billion dense parameters, including a 460 million-parameter vision encoder and a 420 million-parameter multi-token prediction head.
- Architecture: 64 layers, arranged as three Gated DeltaNet linear-attention layers followed by one full-attention layer.
- Inputs and outputs: Text and images in, text out, with reasoning mode enabled by default.
- Context: Up to 262,144 tokens per model call.
- Approximate weight footprint: 52 GB in BF16, 29 GB in FP8, and 23 GB in NVFP4. NVFP4 requires Blackwell hardware.
- Listed hardware: H200, B200, GB300, RTX PRO 6000, RTX 5090, and DGX Spark.
The memory figures describe model weights. Inference also requires memory for the key-value cache, activations, and serving overhead, particularly when requests approach the full context window.