跳到正文

RadixArk 发布 DaisyRose-27B,旨在实现更高效的后训练

Hugging Face Blog当事人·· 22 小时前精选DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
打开原文
原文
图片来自 Hugging Face Blog
谁
RadixArk
引文
Today we are releasing DaisyRose-27B , RadixArk’s first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles .

发生了什么

RadixArk 发布了 DaisyRose-27B,这是其首个模型,基于 Qwen3.8-27B 构建,并使用 Miles 通过强化学习进行训练。它在 144 块 H200 GPU 上,利用 1,200 项精选代理任务进行了 18 天的训练。该模型将 Terminal-Bench 4.1 的 pass@1 从 5.6% 提升至 12.0%,并且在列出的所有基准测试中得分均高于 Qwen3.8-27B。检查点采用 Apache 2.0 协议发布。

推荐理由

DaisyRose-27B 在接受 1,200 项精选任务的训练后,将 Terminal-Bench 4.1 的 pass@1 从 Qwen3.8-27B 的 5.6% 提升至 12.0%,同时其 Apache 2.0 检查点使得该模型可供他人运行。

这件事到目前为止

RadixArk has released DaisyRose-27B, its first AI model, which utilizes reinforcement learning to improve agent task performance over its base, Qwen3.8-27B. The company stated that the model was trained using a method called Miles on 1,200 curated agent tasks over 18 days, utilizing 144 H200 GPUs. According to RadixArk, the model outperforms the base Qwen3.8-27B on every listed benchmark. Specifically, the company reported that DaisyRose-27B increased the Terminal-Bench 4.1 pass@1 rate from 5.6% to 12.0%. The model's checkpoints are available for public use under the Apache 2.0 license, supporting BF16, FP8, and NVFP4 formats.

接下来读什么

查看这件事的完整进展RadixArk Releases DaisyRose-27B, Improving Performance Over Qwen3.8-27B
这条收录于 11月19日 AI 日报看这一期日报的全部内容