THE QUESTION

RadixArk’s DaisyRose‑27B doubles coding benchmark performance while cutting training cost

Can a 27‑billion‑parameter model learn faster and cheaper without sacrificing capability? RadixArk says yes with DaisyRose‑27B, its first model built on the open‑source Qwen3.8‑27B architecture. The team re‑engineered the post‑training pipeline, using a reinforcement‑learning framework called Miles that runs asynchronously while the model serves tasks, aiming to squeeze more learning out of each GPU cycle.

The result is a model that retains the original Qwen3.8‑27B tokenizer, chat template, image understanding, and a massive 256K‑token context window, but promises dramatically lower compute cost. By redesigning the training tasks and backend, DaisyRose‑27B targets the same high‑level capabilities while delivering a leaner, more efficient training experience.

Source: Hugging Face: a]:hidden"> DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency ↗

RadixArk backs its claim with benchmark numbers: Terminal‑Bench 4.1 pass@1 jumps from 5.6 % to 12.0 %, more than doubling the original score. Across a suite of reasoning, instruction‑following, and coding tests the model consistently outperforms its predecessor, with gains on DeepSWE, SWE Bench Pro, IFBench and GPQA Diamond as well.

The efficiency story is equally striking. Training ran on 144 Nvidia H200 GPUs for 18 days, using only 1,200 curated long‑horizon tasks. The asynchronous backend achieved a model‑FLOPs utilization (MFU) of 37 %, translating to a 57 % reduction in GPU time compared with a comparable Megatron‑based trainer. Cloud sandbox costs fell about 80 % thanks to dynamic memory sharing, and the entire run consumed roughly 63,000 GPU‑hours.

Source: Hugging Face: a]:hidden"> DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency ↗

The performance boost comes with practical trade‑offs. DaisyRose‑27B still requires a sizable GPU cluster—144 H200s—for the full training run, which is out of reach for most individual researchers or small startups. The model also inherits the Qwen3.8‑27B architecture, so any limitations of that base (such as token‑level biases) remain present.

Moreover, while the checkpoints are released under Apache 2.0, deploying the model efficiently demands additional tooling like SGLang and DSpark speculative decoding. Teams without expertise in these serving stacks may face a steep integration curve, and the high‑capacity 256K context window can strain memory on modest hardware.

If you’re curious about the efficiency gains, the easiest step is to pull the DaisyRose‑27B checkpoint from Hugging Face and run it through the SGLang serving guide. The guide includes launch configurations for a range of GPU sizes, letting you test the model on your own prompts without rebuilding the training pipeline.

For developers interested in reproducing the cost savings, explore the open‑source Miles framework and the AgentENV sandbox service described in the release. Experimenting with a subset of the 1,200 curated tasks can give a feel for the asynchronous training loop, and sharing results on the model’s Community tab helps the ecosystem refine the approach.

THE BREAKDOWN

DaisyRose‑27B shows that clever task design and an asynchronous RL backend can halve training cost while delivering solid benchmark gains, making large‑scale post‑training more accessible—provided you have the hardware to run it.

Sources & further reading

  1. Hugging Face: a]:hidden"> DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency ↗

Published by The Breakdown. How we report · Our source desk.