RadixArk opens DaisyRose-27B, an agent-tuned Qwen model
RadixArk has published DaisyRose-27B on Hugging Face, the company's first model. It builds on Qwen3.8-27B and was trained with reinforcement learning to handle long-running agent work, with sessions of up to 1M tokens.
RadixArk has published DaisyRose-27B on Hugging Face, the company's first model. It builds on Qwen3.8-27B and was trained with reinforcement learning to handle long-running agent work, with sessions of up to 1M tokens.
What is released
Checkpoints are available in BF16, FP8 and NVFP4 formats under the Apache 2.0 licence. The model keeps the base model's architecture, tokenizer, chat template, image understanding and 256K-token context window. Longer training sessions were stitched together from several trajectories through summarization by the agent harness.
RadixArk recommends serving it with SGLang plus DSpark speculative decoding, and points to an SGLang cookbook with launch settings for different GPUs and checkpoint formats. Once the server is running, it can be called with the standard OpenAI Python client. The model reasons before answering by default, and returns its reasoning and final answer in separate fields.
Reported results
According to RadixArk, DaisyRose-27B beats Qwen3.8-27B on every benchmark it lists:
- Terminal Bench 4.1: 5.6 to 12.0
- Terminal Bench 2.1: 73.0 to 75.5
- DeepSWE v1.1: 42.2 to 46.2
- SWE Bench Pro v2: 91.3 to 94.7
- IFBench: 79.5 to 82.4
- GPQA Diamond: 89.2 to 90.7
- HLE (text-only): 33.9 to 34.3
The company reports the higher of the BF16 and FP8 runs per benchmark, and says the two differ by under one point where both were measured. These are the vendor's own figures; no independent evaluation is cited.
How it was trained
The training set was about 1,200 tasks, built with Eigent AI from thousands of candidates. Each task was redesigned around a mistake the base model makes, requires several dependent steps, and was reviewed so it cannot be passed without solving the actual problem. Each attempt gets up to 1.5 hours in a sandbox.
Training ran for 18 days on 144 H200 GPUs, 128 for running agents and 16 for the trainer, in a fully asynchronous setup. RadixArk says its internal backend reached 37% model FLOPs utilization against 16% for the open-source trainer, which it translates into 57% less GPU time. The run totalled 63,000 GPU-hours.
The sandbox service ran about 944,000 sandboxes between mid-September and mid-November, up to 1,566 at once. RadixArk reports a creation success rate above 99%, and cloud costs about 80% below popular sandbox providers. The training backend and sandbox service figures describe RadixArk's own infrastructure; only the model weights are the released artifact in this announcement.
What to check
RadixArk asks users to report results on their own tasks, including failures, in the model's Community tab. That is the right test: the gains are measured on coding, terminal and reasoning benchmarks, and the source says nothing about sales, support or back-office workloads.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.