What happened
RadixArk has released DaisyRose-27B, its first model, which is derived from Qwen3.8-27B and post-trained with reinforcement learning on long-horizon agent tasks, RadixArk reported. The company said the model lifted pass@1 on Terminal-Bench 4.1 to 12.0% from the base model's 5.6%. It also reported gains on other reasoning, instruction-following and coding benchmarks. BF16, FP8 and NVFP4 checkpoints are available on Hugging Face.
Training relied on roughly 1,200 tasks chosen from thousands of candidates developed with Eigent AI. According to RadixArk, each task was reworked to expose errors the base model made and adjusted to a moderate solve rate. Each was also reviewed so it could not be passed without actually solving it. A single attempt could run for up to 1.5 hours with more than 100 tool calls, and sessions reached as many as 1M tokens.
The run took 18 days on 144 H200 GPUs using Miles, RadixArk's open-source reinforcement learning framework, in an asynchronous setup: 128 GPUs ran agents while 16 handled training. The company said its backend reached 37% model FLOPs utilization versus 16% for a Megatron-based open-source trainer, cutting GPU time by 57%. It said its AgentENV-based sandbox service ran about 944,000 sandboxes.
Context
The model builds on Qwen3.8-27B from the Qwen team. RadixArk credited the Miles and SGLang communities as well as slime, AgentENV and Harbor. It said its training efficiency gains were larger on mixture-of-experts models: 6.2 times on Qwen3.6-35B-A3B and more than 10 times on GLM-5.3-Flash. It also said its sandbox service scaled from one machine to 12 and cost about 80% less than popular cloud sandbox providers.
ブリーフィング
1週間のテックニュースを1通のメールで。
読む価値のある記事を毎週お届けします。配信停止はいつでも可能です。
