Technology
Hacker News

DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

Source Entity

Hacker News

November 20, 2026
DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

RadixArk has released DaisyRose-27B, an AI model built on Qwen3.8-27B that utilizes reinforcement learning to significantly boost performance on long-horizon agent tasks. Achieving a 12% pass rate on Terminal-Bench 4.1, the model demonstrates high efficiency by training on just 1,200 curated tasks.

The Emergence of DaisyRose-27B: A New Efficiency Paradigm

RadixArk has officially introduced DaisyRose-27B, marking the organization's debut model release. Built upon the foundation of Qwen3.8-27B, this model represents a significant shift in how researchers approach post-training optimization. By leveraging reinforcement learning through a framework known as 'Miles,' the developers have demonstrated that targeted, high-quality training can yield substantial performance improvements without requiring the massive, indiscriminate datasets often associated with large language model training.

Architectural Foundation and Contextual Capacity

At its core, DaisyRose-27B retains the robust architecture of the original Qwen3.8-27B, including its tokenizer, chat template, and image understanding capabilities. Crucially, it maintains the 256K-token context window, which is essential for complex, multi-step agentic workflows. By building on an established base, RadixArk has avoided the need to reinvent the fundamental neural architecture, instead focusing their engineering resources on the post-training alignment that dictates how a model interacts with long-horizon tasks.

Breakthroughs in Long-Horizon Agent Tasks

The most notable achievement of this release is the model's performance on long-horizon agent tasks, with sessions scaling up to 1 million tokens. The training process utilized a highly curated set of only 1,200 tasks, executed over 18 days using 144 H200 GPUs. This constrained compute budget highlights a growing trend in the AI industry: the move away from 'bigger is better' toward 'smarter is faster.' By utilizing harness summarization to span multiple trajectories, the model effectively learns to manage extended temporal dependencies that typically baffle standard LLMs.

Benchmarking Success and Quantitative Gains

Quantitatively, the results are striking. DaisyRose-27B improved the Terminal-Bench 4.1 pass@1 rate from 5.6% to 12.0%. This doubling of performance on a specialized benchmark for agentic reasoning underscores the efficacy of the Miles training methodology. Beyond Terminal-Bench, the model shows consistent gains across general reasoning, instruction following, and coding benchmarks, suggesting that the reinforcement learning process has instilled a more generalized capability for logical problem-solving.

Accessibility and Future Implications

In a move to encourage community collaboration, RadixArk has released the checkpoints under the Apache 2.0 license, providing versions in BF16, FP8, and NVFP4 formats. This accessibility is vital for the open-source ecosystem, as it allows researchers to experiment with different precision levels and inference settings. The success of DaisyRose-27B suggests that future AI development will likely prioritize the quality of task curation and the efficiency of the training loop, potentially lowering the barrier to entry for smaller labs to produce models that rival those created by major industry players.

Multiple Citing Sources

Verification Required?

Read the full report from the primary source

Go to Hacker News