What was released and how it was trained

RadixArk introduced DaisyRose-27B in a Hugging Face post, which adapts its original article. It is the company's first model. It is based on Qwen3.8-27B and was trained with reinforcement learning through Miles, the company's open framework for post-training large models. The checkpoints are available under the Apache 2.0 license in BF16, FP8 and NVFP4.

The model keeps the architecture, tokenizer, chat template, image understanding and 256K-token context window of Qwen3.8-27B. The longest training sessions cover multiple trajectories with harness summarization. The target was long-horizon agent tasks, with sessions of up to 1M tokens.

The data came from a collaboration with Eigent AI. From a large pool of tasks in machine learning, software engineering and science, the two teams built thousands of candidate tasks and kept about 1,200. RadixArk describes three selection criteria. Each task targets a mistake the base model makes, and the solve rate was tuned so that it is useful for learning. Each task requires interdependent steps, with up to 1.5 hours in a sandbox, more than 100 tool calls and more than 256K output tokens per attempt. Finally, the tasks must reward genuine solutions: those that passed without solving the problem, or had a design flaw, were fixed or removed.

Benchmark results: the company's own measurements

The results below come from RadixArk itself, for its own model. They are not an independent evaluation. The company says DaisyRose-27B outperforms the base Qwen3.8-27B on all benchmarks in the table. The largest gain is on Terminal Bench 4.1, where pass@1 goes from 5.6% to 12.0%. On HLE (text-only) the difference is the smallest, from 33.9 to 34.3.

On methodology, the company says that for each benchmark the highest score between the BF16 and FP8 runs is recorded. Where both formats were evaluated, their differences were less than one point. According to the post, the comparisons and evaluation settings are in the model card.

RadixArk also asks users to try the model on their own tasks and to record their results, and the areas where it falls short, on the model's Community tab.

Qwen3.8-27B and DaisyRose-27B scores as published by RadixArk (highest BF16/FP8 score per benchmark)

BenchmarkQwen3.8-27BDaisyRose-27B
Terminal Bench 4.15.612.0
Terminal Bench 2.173.075.5
DeepSWE v1.142.246.2
SWE Bench Pro v291.394.7
HLE (text-only)33.934.3
IFBench79.582.4
GPQA Diamond89.290.7

Training backend and sandboxes: infrastructure measurements

Training ran in fully asynchronous mode: the agent keeps working while the model learns. Of the 144 H200 GPUs, 128 run inference for the agent tasks and 16 run the trainer. After each step, the updated model is distributed to the 128 GPUs. Learning runs on the company's in-house backend, designed for agent training, where each step learns from a few very long sessions.

According to RadixArk, the backend achieved 37% model FLOPs utilization (MFU, the share of peak compute that goes into useful training work), while the open Megatron-based trainer reached 16% in comparable runs. The company translates this into 57% less GPU time for the same training. For mixture-of-experts models it reports a larger gain: 6.2× on Qwen3.6-35B-A3B and more than 10× on GLM-5.3-Flash, of about 300B parameters.

Counting also the time the agent spends on its tasks, the release run processed about 16,800 tokens per second on the 144 GPUs, versus about 8,300 for the open trainer on the same GPUs. The run lasted weeks and 63,000 GPU hours in total, with steady speed and performance from the first step to the last.

Each training attempt runs in a sandbox, that is, a fresh copy of the task environment where the agent executes commands, edits files and is scored. The service is based on the open project AgentENV. From mid-September to mid-November it ran about 944,000 sandboxes, with up to 1,566 concurrently. It grew from one machine to up to 12 and received ten upgrades, almost all while training was in progress. Over the nine weeks, 99% of sandbox creations succeeded, and since late October the rate has exceeded 99.95%.

The company says the cloud infrastructure cost for the sandboxes was about 80% lower than running the same sandboxes on popular providers. One of the ways is that each sandbox's memory scales up and down with usage, so that more sandboxes share the same machine. For serving the model it recommends SGLang with DSpark speculative decoding.