Oossa

RadixArk releases DaisyRose-27B, an RL‑only coding model

The 27‑billion‑parameter model improves Terminal Bench 4.1 pass@1 to 12 % and runs on a single GPU; weights are open‑source under Apache‑2.0.

By Published by Oossa: 1 min read

Fatos Bytyqi · Unsplash

RadixArk announced the launch of DaisyRose-27B on 20 November 2026. The model is built on Qwen3.8-27B and was fine‑tuned only with reinforcement learning (RL), without any supervised data. RadixArk says the new model more than doubles the pass@1 score on the long‑horizon Terminal Bench 4.1 benchmark, reaching 12.0 % compared with the base model’s 5.6 %. The company also released BF16, FP8 and NVFP4 checkpoints on Hugging Face, and the model can be run on a single GPU ranging from data‑center H200 cards to consumer RTX 5090 GPUs.

How DaisyRose was trained

Training used RadixArk’s open‑source Miles framework on 144 H200 GPUs for 18 days, totaling about 63,000 GPU‑hours. The team curated 1,200 long‑horizon tasks in partnership with Eigent AI, each designed to expose mistakes made by the base model. Sessions could be up to 1 million tokens and lasted as long as 1.5 hours, with more than 100 tool calls per attempt. The RL loop ran asynchronously: 128 GPUs handled the agent’s actions while 16 GPUs performed the learning updates, keeping the model’s compute utilization 57 % higher than a comparable open‑source trainer. Over the nine‑week run, the sandbox service created roughly 944,000 sandbox instances, with a success rate above 99 %.

Performance on benchmarks

Besides Terminal Bench 4.1, DaisyRose-27B scores higher than Qwen3.8-27B on every listed benchmark. On Terminal Bench 2.1 it reaches 75.5 % versus 73.0 % for the base model. DeepSWE v1.1 improves from 42.2 % to 46.2 %, and SWE Bench Pro v2 rises to 94.7 % from 91.3 %. The model also beats larger open models on IFBench (82.4 % vs 79.5 % for Qwen3.8‑27B) and matches GPT‑6 Luna on GPQA Diamond (90.7 % vs 89.2 %). All scores are the higher of the BF16 and FP8 runs, which differ by less than one point.

Why it matters

Developers can now try a state‑of‑the‑art coding assistant without paying for a cloud API, since the model runs on a single GPU and is free to download. The RL‑only training shows that significant gains are possible without large supervised datasets, which may lower the barrier for smaller teams to build specialized agents.

Was this article useful?
Share

Read next

Oossa · Newsletter

The week in AI, explained

Every Monday: the stories worth knowing, in plain language. Free, no spam.