Models·
RadixArk's DaisyRose-27B Doubles Agent Benchmark Scores with Efficient RL Post-Training
RadixArk releases DaisyRose-27B, a Qwen3.8-27B derivative trained with reinforcement learning on 144 H200 GPUs. The model more than doubles Terminal-Bench 4.1 pass@1 to 12% using only 1,200 curated long-horizon tasks, demonstrating how targeted post-training can yield outsized gains for agent workloads.

RadixArk has debuted its first model, DaisyRose-27B, built atop Qwen3.8-27B and refined through reinforcement learning using the Miles framework. The training run consumed 18 days on a cluster of 144 NVIDIA H200 GPUs — a modest footprint for a 27-billion-parameter model — yet delivered a striking leap on agent-centric benchmarks. Terminal-Bench 4.1 pass@1 climbed from 5.6% to 12.0%, while gains appeared across coding, reasoning, and instruction-following suites including SWE-Bench Pro, DeepSWE, and GPQA Diamond.
Why it matters for GPU / AI infrastructure
The project underscores a growing trend: high-impact post-training no longer demands massive GPU-months. By curating just 1,200 long-horizon tasks — each spanning up to 1 million tokens and 1.5 hours of sandbox execution — RadixArk and partner Eigent AI concentrated compute on the failure modes that mattered most. The result is a model that thinks before answering, emits reasoning traces, and runs efficiently under SGLang with DSpark speculative decoding on BF16, FP8, or NVFP4 checkpoints.
For infrastructure teams, the takeaway is clear: workload-aware task design and efficient RL backends can extract more capability per GPU-hour than scaling pre-training alone. DaisyRose-27B retains the base model's 256K context window, image understanding, and Apache 2.0 licensing, making it a drop-in upgrade for agent pipelines that need stronger planning and tool-use without expanding the hardware budget.
- aigpu
- ai gpu
- ai gpu cloud
- aigpu dubai
- reinforcement-learning
- h200
- post-training
- agent-tasks
By AiGpu Editorial · Editorial rewrite based on public reporting (Hugging Face Blog)
← All articles