初创公司 RadixArk 发布了基于 Qwen3.8-27B 构建的首款大语言模型 DaisyRose-27B 1。该模型通过仅使用约 1,200 个精心设计的长程智能体任务进行为期 18 天的强化学习训练,将 Terminal-Bench 4.1 基准测试中的 pass@1 准确率从 5.6% 提升至 12.0% 1。此次训练使用了 144 块 H200 GPU,总计消耗 63,000 GPU-hours 1。
此次发布展示了在受限算力预算下,高效基础设施和高质量数据对提升模型 Agent 能力的重要性 1。RadixArk 采用的 Miles 框架相比开源 Megatron 后端展现出显著的效率优势,其 MFU 达到 37%(而后者为 16%),使训练成本降低了 57%,端到端速度加快了两倍 1。此外,该基础设施运行了约 944,000 个沙箱实例,相关成本比主流云服务商低约 80% 1。目前,DaisyRose-27B 已提供 BF16、FP8 和 NVFP4 多种精度的检查点供下载 1。
Startup RadixArk has released its first large language model, DaisyRose-27B, which is built upon the Qwen3.8-27B architecture 1. The model demonstrates significant improvements in agent capabilities through a targeted reinforcement learning process using approximately 1,200 carefully curated long-horizon tasks 1. This focused training approach elevated the pass@1 accuracy on Terminal-Bench 4.1 from 5.6% to 12.0% 1.
The development of DaisyRose-27B relied on 144 H200 GPUs over an 18-day period, totaling 63,000 GPU-hours 1. Utilizing the Miles framework infrastructure, the training achieved a Model FLOPs Utilization (MFU) of 37%, compared to 16% for open-source Megatron backends, resulting in a 57% reduction in training costs and a twofold increase in end-to-end speed 1. During the process, about 944,000 sandbox instances were executed, with infrastructure costs approximately 80% lower than those of mainstream cloud providers 1. Checkpoints in BF16, FP8, and NVFP4 formats are available for download 1.
评论
还没有评论,欢迎留下第一条。