The development

What happened

RadixArk has released DaisyRose-27B, a Qwen3.8-27B derivative trained with reinforcement learning on about 1,200 curated agent tasks. The developer says it trained for 18 days on 144 H200 GPUs and retains the base model's 256K-token context and image understanding. The release's benchmark comparisons need qualification: a prominent Terminal Bench comparison mixes versions 4.0 and 4.1.

Key facts

  • RadixArk announces BF16, FP8 and NVFP4 checkpoints under Apache 2.0. Its model card describes reinforcement-learning-only post-training, retaining the base model's architecture, tokenizer and chat template.

  • Training sessions could span up to 1M tokens through harness summarization across multiple trajectories. That is not a 1M-token inference context: the model card lists a 262,144-token context window.

  • The company says it selected tasks around mistakes made by the base model, used multi-step sandboxed jobs and reviewed attempts to remove flawed or easily gamed rewards. These are developer-described training methods, not an independent assessment of reliability.

  • RadixArk reports 12.0% on Terminal Bench 4.1, but the model card identifies its 5.6% Qwen3.8-27B comparator as Terminal Bench 4.0. This is not a controlled same-version doubling.

  • The published DaisyRose scores take the higher of BF16 and FP8 runs. Other model comparisons use scores from different developer sources and harnesses, so the table should not be read as a uniform independent ranking.

The context

Why it matters

The release targets multi-step agents by selecting difficult tasks and grading actual task completion, not simply making sessions longer. Training-session length, live context and benchmark settings remain different measures. DaisyRose's Terminal Bench 4.1 result uses eight repetitions per task in an internal mini-swe-agent harness with summarization. Reported gains warrant task-specific evaluation, not claims of universal superiority or lower deployment cost; TideReach has not independently reproduced them.