RadixArk发布DaisyRose-27B:纯RL微调的开源编程Agent
RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only
纯RL后训练且无需SFT的编程Agent模型,Terminal Bench成绩翻倍,为探索低成本高效能Agent训练路径提供了极具参考价值的一手实践数据。
This week RadixArk released DaisyRose-27B, its first open-weight model. DaisyRose-27B is an agentic coding model from RadixArk, post-trained from Qwen3.8-27B with reinforcement learning (RL). The model card page states that it used RL only, with no SFT or distillation. It ships under Apache 2.0 as a drop-in replacement for the base model. The main result: Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
本周,RadixArk 发布了 DaisyRose-27B,这是其首款开放权重模型。DaisyRose-27B 是 RadixArk 推出的一款智能体编码模型,基于 Qwen3.8-27B 使用强化学习(RL)进行后训练。模型卡片页面指出,该模型仅使用了 RL,未进行监督微调(SFT)或蒸馏。它以 Apache 2.0 许可证发布,可作为基础模型的即插即用替代品。主要成果:Terminal Bench 4.1 的 pass@1 指标从 5.6 提升至 12.0。你可以通过我们的 GPULeaderboard.com Chrome 扩展程序在此查看部署选项和规格。
TL;DR
TL;DR(太长不看版)
- Size: 27.8B dense parameters (0.46B vision encoder, 0.42B MTP head), 262,144-token context.
- Runs on: 1 GPU. BF16 is about 52 GB, FP8 about 29 GB, NVFP4 about 23 GB (Blackwell only). Listed targets: H200, B200, GB300, RTX PRO 6000, RTX 5090, DGX Spark. GGUF also available.
- Performance: Beats its Qwen3.8-27B base on all 7 reported benchmarks.
- Best: IFBench 82.4, the top score among the 9 models RadixArk compared.
- Worst: HLE (text only) 34.3, only +0.4 over the base and below Inkling 2 Small’s 44.1.
- Bottom line: Best: Terminal Bench 4.1 more than doubles after 18 days of RL. Worst: it still trails much larger MoE models on hard agentic tasks.
- 规模:278 亿密集参数(其中 4.6 亿为视觉编码器,4.2 亿为 MTP 头),上下文窗口为 262,144 token。
- 运行环境:1 张 GPU。BF16 精度约需 52 GB 显存,FP8 约需 29 GB,NVFP4 约需 23 GB(仅限 Blackwell 架构)。官方列出的目标硬件包括:H200、B200、GB300、RTX PRO 6000、RTX 5090、DGX Spark。同时也提供 GGUF 格式。
- 性能:在所有 7 项报告的基准测试中均优于其 Qwen3.8-27B 基础模型。
- 最佳表现:IFBench 得分 82.4,在 RadixArk 对比的 9 款模型中位居榜首。
- 最差表现:HLE(纯文本)得分为 34.3,仅比基础模型高出 0.4,且低于 Inkling 2 Small 的 44.1。
- 总结:最佳方面:经过 18 天的 RL 训练后,Terminal Bench 4.1 成绩翻倍以上。最差方面:在复杂的智能体任务上,它仍落后于参数量更大的 MoE 模型。
What is DaisyRose-27B?
什么是 DaisyRose-27B?
DaisyRose-27B keeps the Qwen3.8-27B architecture unchanged. The model card lists 64 decoder layers in a 3:1 pattern. Three Gated DeltaNet linear-attention layers alternate with 1 full-attention layer. Hidden size is 5,120. A 27-layer vision encoder handles images. One multi-token-prediction layer supports speculative decoding. Input is text and images, output is text, and it is a thinking model.
DaisyRose-27B 保持了 Qwen3.8-27B 的架构不变。模型卡片列出其包含 64 层解码器,采用 3:1 的模式排列。三个门控 DeltaNet 线性注意力层与一个全注意力层交替出现。隐藏层大小为 5,120。配备一个 27 层的视觉编码器用于处理图像。一个多令牌预测层支持推测性解码。输入为文本和图像,输出为文本,并且是一款具备思考能力的模型。
How was DaisyRose-27B trained?
DaisyRose-27B 是如何训练的?
RadixArk built about 1,200 curated tasks with Eigent AI, chosen from thousands of candidates. Each task targets a specific failure of the base model. Each sandbox attempt allows up to 1.5 hours, 100+ tool calls and 256k+ output tokens. Agent sessions reach up to 1M tokens, per the blog.
RadixArk 利用 Eigent AI 构建了约 1,200 个精心策划的任务,这些任务是从数千个候选者中筛选出来的。每个任务都针对基础模型的特定缺陷。每次沙盒尝试最多允许 1.5 小时、100+ 次工具调用和 256k+ 输出 token。据博客文章所述,智能体会话最多可达 1M token。
Training ran on Miles, RadixArk’s open-source RL framework, in fully asynchronous mode. The run used 144 H200 GPUs for 18 days. Of those, 128 handled agent inference and 16 handled training. Total compute was 63,000 GPU-hours.
训练在 Miles 上进行,这是 RadixArk 的开源 RL 框架,采用完全异步模式。此次运行使用了 144 张 H200 GPU,历时 18 天。其中 128 张用于智能体推理,16 张用于训练。总计算量为 63,000 GPU 小时。
RadixArk reports 37% MFU on its internal backend. An open-source Megatron-based trainer reached 16%. End-to-end throughput was 16.8k tokens/s versus 8.3k, a 2.0x speedup. RadixArk says this cut GPU time by 57%.
RadixArk 报告其内部后端实现了 37% 的 MFU(有效浮点利用率)。一个基于开源 Megatron 的训练器达到了 16%。端到端吞吐量为每秒 16.8k token,而对照组为 8.3k,速度提升了 2.0 倍。RadixArk 表示这使 GPU 使用时间减少了 57%。
Sandboxes ran on AgentENV. RadixArk reports about 944,000 sandboxes from mid-September to mid-November. Peak concurrency hit 1,566. Creation success was 99% over 9 weeks. RadixArk puts the cost about 80% below popular cloud sandbox providers.
沙箱运行在 AgentENV 上。RadixArk 报告称,从9月中旬到11月中旬共有约944,000个沙箱。峰值并发数达到1,566。在9周内创建成功率为99%。RadixArk 指出其成本比主流云沙箱提供商低约80%。
How does DaisyRose-27B perform on benchmarks?
DaisyRose-27B 在基准测试中的表现如何?
All scores are self-reported by RadixArk. DaisyRose scores are the higher of its BF16 and FP8 runs.
所有分数均由 RadixArk 自行报告。DaisyRose 的分数取其 BF16 和 FP8 运行结果中的较高者。
| Benchmark | DaisyRose-27B | Qwen3.8-27B | Change |
|---|---|---|---|
| Terminal Bench 4.1 (pass@1) | 12.0 | 5.6 | +6.4 |
| Terminal Bench 2.1 | 75.5 | 73.0 | +2.5 |
| DeepSWE v1.1 | 46.2 | 42.2 | +4.0 |
| SWE Bench Pro v2 | 94.7 | 91.3 | +3.4 |
| HLE (text only) | 34.3 | 33.9 | +0.4 |
| IFBench | 82.4 | 79.5 | +2.9 |
| GPQA Diamond | 90.7 | 89.2 | +1.5 |
| 基准测试 | DaisyRose-27B | Qwen3.8-27B | 变化 |
|---|---|---|---|
| Terminal Bench 4.1 (pass@1) | 12.0 | 5.6 | +6.4 |
| Terminal Bench 2.1 | 75.5 | 73.0 | +2.5 |
| DeepSWE v1.1 | 46.2 | 42.2 | +4.0 |
| SWE Bench Pro v2 | 94.7 | 91.3 | +3.4 |
| HLE (仅文本) | 34.3 | 33.9 | +0.4 |
| IFBench | 82.4 | 79.5 | +2.9 |
| GPQA Diamond | 90.7 | 89.2 | +1.5 |
Source: RadixArk blog and model card.
来源:RadixArk 博客及模型卡片。
The largest jump is Terminal Bench 4.1, which more than doubles. Coding gains are steady: +4.0 on DeepSWE v1.1 and +3.4 on SWE Bench Pro v2. Knowledge gains are small, at +0.4 on HLE and +1.5 on GPQA Diamond.
增幅最大的是 Terminal Bench 4.1,成绩翻倍有余。编码能力稳步提升:DeepSWE v1.1 增加4.0,SWE Bench Pro v2 增加3.4。知识类能力提升较小,HLE 增加0.4,GPQA Diamond 增加1.5。
Against larger models, results are mixed. MiMo V2.6 Flash (309B total) leads Terminal Bench 4.1 at 16.1. Inkling 2 Small (594B total) leads DeepSWE at 72.1 and HLE at 44.1. DaisyRose-27B tops RadixArk’s 9-model table only on IFBench.
与更大规模的模型相比,结果喜忧参半。MiMo V2.6 Flash(总参数309B)在 Terminal Bench 4.1 中以16.1领先。Inkling 2 Small(总参数594B)在 DeepSWE 中以72.1领先,在 HLE 中以44.1领先。DaisyRose-27B 仅在 IFBench 一项上位居 RadixArk 九模型榜单首位。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力