RadixArk releases DaisyRose-27B, a reinforcement-learning-tuned agent model
RadixArk's first model, built on Qwen3.8-27B, was trained with reinforcement learning on long-horizon agent tasks and released under Apache 2.0.
In plain words
RadixArk trained an existing open model, Qwen3.8-27B, with reinforcement learning so it handles long, multi-step agent jobs better. The company says the tuned model, DaisyRose-27B, scores higher on the benchmarks it ran, with the biggest jump on Terminal Bench 4.1. The checkpoints are downloadable under the Apache 2.0 license.
What was announced
DaisyRose-27B keeps the base model's architecture, tokenizer, chat template, image understanding, and 256K-token context window. Training used about 1,200 curated tasks, built with Eigent AI and selected because the base model made mistakes on them, with sessions up to 1M tokens. The run took 18 days on 144 H200 GPUs using RadixArk's Miles reinforcement learning framework, with attempts running in sandboxes built on AgentENV.
RadixArk reports that Terminal Bench 4.1 pass@1 rose from 5.6% to 12.0%, the largest gain among the benchmarks listed, and that DaisyRose-27B beat Qwen3.8-27B on all of them, including SWE Bench Pro v2 and GPQA Diamond. Scores use the higher of BF16 and FP8 runs, which differed by less than one point. The company also claims 57% lower training cost than an open-source Megatron-based trainer and about 16,800 tokens per second across the 144 GPUs, but these efficiency figures come from RadixArk's own comparisons and are not independently confirmed.
What it does not show
- All results and efficiency comparisons are reported by RadixArk in its own announcement; no independent evaluation is cited.
- Benchmark scores use the higher of BF16 and FP8 runs, so reported figures may not reflect a single checkpoint configuration.
For security and IT teams
- Access
- Checkpoints are available under Apache 2.0 in BF16, FP8, and NVFP4 formats, and RadixArk recommends serving with SGLang and DSpark speculative decoding.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.