hn.today

DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

radixark.com157 points45 comments
Screenshot of DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

DaisyRose-27B is RadixArk’s reinforcement-learning post-trained version of Qwen3.8-27B, developed to improve long-horizon agent work without relying on a large task set. Its strongest reported result is a rise in Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%; it also exceeds the base model on six other benchmarks, including SWE-Bench Pro v2 (94.7 versus 91.3), IFBench (82.4 versus 79.5), and GPQA Diamond (90.7 versus 89.2). Training used about 1,200 tasks selected from thousands of candidates with Eigent AI. Each task was designed to expose a specific weakness in the base model, require a chain of reasoning and tool use, and resist shortcuts or reward hacking. Agents could spend up to 90 minutes per attempt, with more than 100 tool calls and 256,000 output tokens.

The team argues that careful task design was enabled by efficient training and dependable execution infrastructure. Over 18 days, 144 H200 GPUs trained the model using Miles in an asynchronous setup: 128 GPUs ran agents and 16 updated the model. RadixArk reports 37% model FLOPs utilization, compared with 16% for an open-source trainer, and twice the end-to-end throughput - 16,800 versus 8,300 tokens per second. Its sandbox service handled about 944,000 task environments, reaching 1,566 concurrent instances; creation success exceeded 99.95% late in the run, while sandbox costs were about 80% below those of popular providers. BF16, FP8, and NVFP4 checkpoints are available, along with the Miles framework for training agents.

Read on radixark.com45 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.