Grafy News
Try Grafy free

DaisyRose-27B Sets New Efficiency Bar for Post‑Training LLMs

RadixArk released DaisyRose-27B, a Qwen‑based model that doubles Terminal‑Bench pass@1 while cutting training cost by 57%.

Illustration for the news story: DaisyRose-27B Sets New Efficiency Bar for Post‑Training LLMsSource: Hugging Face Blog

DaisyRose-27B, the first model from RadixArk built on Qwen3.8‑27B, was released on November 18, 2026, and immediately demonstrated a notable jump in both performance and training efficiency. The model’s improvements matter because they show how targeted task design and an asynchronous reinforcement‑learning pipeline can halve the cost of large‑scale post‑training while delivering stronger results on a suite of reasoning and coding benchmarks.

What is DaisyRose-27B and why it matters?

According to the Hugging Face Blog, DaisyRose-27B inherits the 27‑billion‑parameter architecture, tokenizer, chat template, image‑understanding capabilities, and a 256K‑token context window of its predecessor Qwen3.8‑27B. The key difference lies in a two‑week reinforcement‑learning (RL) run that exposed the model to long‑horizon agent tasks up to 1 million tokens per session. By focusing on 1,200 carefully curated tasks, the team achieved a pass@1 increase on Terminal‑Bench 4.1 from 5.6% to 12.0% – more than double the original score. The release is offered under an Apache 2.0 license in BF16, FP8, and NVFP4 checkpoint formats, making it accessible for both research and production.

How does its performance compare to Qwen3.8‑27B?

The blog provides a side‑by‑side benchmark table. Across all listed tests – Terminal‑Bench 4.1, Terminal‑Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, HLE (text‑only), IFBench, and GPQA Diamond – DaisyRose‑27B scores higher than the base Qwen3.8‑27B. The most dramatic lift appears on Terminal‑Bench 4.1, where pass@1 rises from 5.6% to 12.0%. Other gains are modest but consistent, e.g., IFBench improves from 79.5% to 82.4% and GPQA Diamond from 89.2% to 90.7%. Scores were taken as the higher of the BF16 and FP8 runs, which differed by less than one point.

What training techniques enabled the efficiency gains?

Three pillars underpinned the release:

  1. High‑quality training data – In partnership with Eigent AI, RadixArk filtered a large pool of machine‑learning, software‑engineering, and scientific tasks down to 1,200 that specifically targeted mistakes made by Qwen3.8‑27B. Each task required multi‑step reasoning, planning, and execution, often involving more than 100 tool calls and generating up to 256K output tokens.
  2. Efficient asynchronous RL backend – The open‑source Miles framework runs the agent and the trainer on separate GPU groups (128 inference GPUs and 16 trainer GPUs) so neither waits for the other. This pipeline achieved 57% lower training cost, with 37% of each GPU’s peak compute dedicated to useful model FLOPs (MFU). By contrast, a comparable open‑source trainer built on Megatron reached only 16% MFU.
  3. Robust sandbox service – Using the AgentENV project, each training attempt ran in an isolated sandbox that could execute commands, edit files, and be graded. Over the nine‑week run, 944,000 sandboxes were launched, 99% of which succeeded, and sandbox cost was about 80% lower than using commercial providers.

The combined effect was a 2× faster end‑to‑end training speed (≈16,800 tokens / s across 144 H200 GPUs) and stable performance over 63,000 GPU‑hours of compute.

How can developers and researchers use DaisyRose-27B today?

The model can be downloaded directly from Hugging Face and served with SGLang’s speculative decoding for low‑latency chat. The blog includes a quick‑start snippet using the OpenAI Python client:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="none")
response = client.chat.completions.create(
    model="DaisyRose-27B",
    messages=[{"role": "user", "content": "How many positive divisors does 360 have?"}],
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20, "min_p": 0.0},
)
print(response.choices[0].message.reasoning_content)
print(response.choices[0].message.content)

The model returns both its chain‑of‑thought reasoning (in reasoning_content) and the final answer (in content). For those who prefer a no‑code path, the model can be loaded into any platform that supports OpenAI‑compatible endpoints, including Grafy’s own Grafy Code environment, which lets you spin up a container, install SGLang, and run the snippet with a single click.

Researchers interested in reproducing the training pipeline can clone the Miles repository, follow the provided SGLang serving guide, and adapt the task generation scripts that RadixArk released alongside the model card. Because the checkpoints are Apache‑2.0, they can be fine‑tuned further on domain‑specific data or integrated into multi‑model ensembles.

Grafy relevance: tools to work with DaisyRose-27B

Grafy users have three immediate ways to experiment with DaisyRose-27B:

  • Grafy Code – Deploy the model locally with the SGLang cookbook, edit the Python client script, and iterate on prompt engineering or tool‑use strategies.
  • Grafy Chat – Connect to the locally served endpoint via the multi‑model chat UI, compare DaisyRose‑27B’s answers side‑by‑side with other open models, and capture the reasoning traces for analysis.
  • Grafy Reader – Load the model card PDF into the Reader app to search for detailed evaluation settings, hyper‑parameters, and the full list of curated tasks.

These apps let creators, developers, and researchers explore the model’s capabilities without leaving the Grafy workspace, turning a research release into a hands‑on playground.


Source: Hugging Face Blog (https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin)

Frequently asked questions

How much faster was DaisyRose-27B’s training compared to a standard Megatron trainer?

The asynchronous Miles backend processed about 16,800 tokens per second on 144 H200 GPUs, roughly twice the 8,300 tokens per second achieved by a comparable Megatron‑based trainer.

Does DaisyRose-27B retain Qwen3.8-27B’s image‑understanding abilities?

Yes. The model keeps the original architecture’s image‑understanding module and can handle multimodal inputs alongside its extended 256K‑token context window.

Can I fine‑tune DaisyRose-27B on my own dataset?

Because the checkpoints are released under Apache 2.0, you may fine‑tune them with any compatible framework (e.g., Miles, LoRA, or standard PyTorch) as long as you respect the license terms.


Source: Hugging Face Blog — DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency