RadixArk · 2026-11-18 · notable
DaisyRose-27B — RadixArk's open RL-tuned Qwen3.8-27B for long agent tasks
In short
DaisyRose-27B is RadixArk's first model, an Apache-2.0 version of Qwen3.8-27B trained with reinforcement learning on long agent tasks. It beats the base model on every benchmark RadixArk lists, from 5.6 to 12.0 on Terminal Bench 4.1.
RadixArk · 2026-11-18 · 3 sources checked

RadixArk's first open model shows what its RL training stack adds to Qwen3.8-27B on long agent tasks.
Key specs
| Terminal bench 4.1 | 12.0 (base 5.6) |
|---|---|
| Training compute | 63,000 H200 GPU-hours |
What is it?
DaisyRose-27B is an open-weight model from RadixArk, the team behind the Miles reinforcement learning framework. It starts from Qwen3.8-27B and is trained with RL on long-horizon agent tasks, with sessions of up to 1M tokens. The weights are on Hugging Face under Apache-2.0 in BF16, FP8 and NVFP4 versions, with a 262,144-token context.
How does it work?
RadixArk trained the model with Miles on about 1,200 curated tasks built with Eigent AI. Training ran for 18 days on 144 H200 GPUs, 128 of them running the agent and 16 running the trainer, for 63,000 GPU-hours. RadixArk says its internal backend reached 37% model FLOPs utilization against 16% for the open-source trainer, and that its sandbox service ran about 944,000 sandboxes for the agent tasks.
Why does it matter?
The biggest gain is on Terminal Bench 4.1, a hard long-horizon benchmark, where the score rises from 5.6 to 12.0. Smaller gains show on DeepSWE v1.1 (42.2 to 46.2) and IFBench (79.5 to 82.4). The release is also a public recipe: it shows how much an RL run of this size can add to a strong open base model.
Who is it for?
teams running open models for coding and terminal agents
Try it
python3 -m sglang.launch_server --model-path RadixArk/DaisyRose-27B