RadixArk Releases DaisyRose-27B: A 27B Agentic Coding Model

RadixArk Releases DaisyRose-27B: A 27B Agentic Coding Model Trained Exclusively With Reinforcement Learning


By Asif Razzaq β€” November 20, 2026


This week, RadixArk released DaisyRose-27B, its first open-weight model. DaisyRose-27B is an agentic coding model post-trained from Qwen3.8-27B using reinforcement learning (RL). According to the model card, training relied on RL alone, with no supervised fine-tuning (SFT) or distillation. The model ships under Apache 2.0 as a drop-in replacement for the base model. The headline result: Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0. Deployment options and specs are available through the GPULeaderboard.com Chrome extension.


TL;DR


  • Size: 27.8B dense parameters (0.46B vision encoder, 0.42B MTP head), 262,144-token context.
  • Runs on: A single GPU. BF16 is about 52 GB, FP8 about 29 GB, and NVFP4 about 23 GB (Blackwell only). Listed targets include H200, B200, GB300, RTX PRO 6000, RTX 5090, and DGX Spark. GGUF is also available.
  • Performance: Beats its Qwen3.8-27B base model on agentic coding benchmarks.

Why the RL-Only Approach Stands Out in 2026


In a landscape where most post-trained coding models rely on a mix of SFT, distillation, and RL, RadixArk's decision to use RL exclusively is notable. It signals growing confidence that well-designed reward signals over agentic tasks can shape capable coding behavior without curated demonstrations. As open-weight agentic models proliferate in 2026, this result offers a useful data point on how far pure RL post-training can go on a mid-sized dense backbone.

via MarkTechPost

Related