Skip to main content
Back to timeline
Hugging Face BlogSource publication:

DaisyRose-27B trains on 1,200 tasks in 18 days and lifts Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%

Synopsis

RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on roughly 1,200 curated long-horizon agent tasks in 18 days on 144 H200 GPUs. Terminal-Bench 4.1 pass@1 rises from 5.6% to 12.0%, and the model also gains on the other reasoning, instruction-following and coding benchmarks listed. The company reports that its internal training backend put 37% of each GPU's peak compute into useful training work, against 16% for an open-source Megatron-based trainer on comparable runs, so the same training needs 57% less GPU time.

AI-generated editorial illustration: DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

Interpretation

DaisyRose-27B is RadixArk's first model, built on Qwen3.8-27B and trained with reinforcement learning on long-horizon agent tasks whose sessions run up to 1M tokens. The training set holds only about 1,200 curated tasks, and the release run finished in 18 days on 144 H200 GPUs. On Terminal-Bench 4.1, pass@1 rises from 5.6% for Qwen3.8-27B to 12.0% for DaisyRose-27B, the largest gain among the benchmarks the company lists.

The tasks were built with Eigent AI, starting from its large pool of tasks in machine learning, software engineering, science and other fields, and about 1,200 of several thousand candidates entered the final training set. Selection used three criteria: each task targets a mistake the original model makes, each requires several interdependent steps, and each must reward actually solving the problem. Each attempt gets up to 1.5 hours in a sandbox with more than 100 tool calls and more than 256K output tokens; the company reviewed every task alongside the model's attempts and fixed or removed tasks that could be passed without solving the intended problem or that had a design flaw.

Training used Miles, RadixArk's open-source reinforcement learning framework, in fully asynchronous mode, so the agent keeps working while the model learns and neither waits for the other. 128 H200s run inference on agent tasks while 16 more run the trainer and learn from the trajectories, and after every step the updated model goes out to the 128 inference GPUs. The company reports 37% model FLOPs utilization on its internal backend for this release against 16% for the open-source Megatron-based trainer on comparable runs, a 57% cut in GPU time for the same training; end to end, counting agent task time, the release run trained at about 16,800 tokens per second across 144 GPUs against about 8,300 for the open-source trainer on the same GPUs.

Every training attempt runs in a sandbox, a fresh copy of the task's environment where the agent runs commands, edits files and is then graded. The service is built on the open-source AgentENV project, to which the company also contributes fixes. Between mid-September and mid-November it ran about 944,000 sandboxes with up to 1,566 running concurrently; the service grew from one machine to as many as 12, shipped ten upgrades nearly all while training was running, and 99% of sandbox creations succeeded across all nine weeks, a figure above 99.95% since late October.

Perspective

The demonstrated scope is a 27B-class model built on Qwen3.8-27B, trained for 18 days on about 1,200 long-horizon agent tasks, with evaluation on Terminal-Bench 4.1, Terminal-Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, HLE (text-only), IFBench and GPQA Diamond. The company recommends serving the model with SGLang and DSpark speculative decoding and points to SGLang cookbook launch configurations for different GPUs and checkpoint formats.

The company states that benchmark scores use the higher of the BF16 and FP8 runs, which differ by less than one point wherever both were evaluated, and that comparison sources and evaluation settings are in the model card. On training efficiency, the 37% MFU and the 57% GPU-time saving come from comparing the company's internal backend with an open-source Megatron-based trainer on comparable runs; the 6.2× and more-than-10× gains on mixture-of-experts models belong to Qwen3.6-35B-A3B and the roughly 300B-parameter GLM-5.3-Flash, not to DaisyRose-27B itself.