Open-Source AI Gets Major Boost: New Model Doubles Performance on Hardest Coding Tasks
Original headline: DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1 — radixark.com
More in AI and machine learning news
RadixArk has released DaisyRose-27B, a new open-source AI model built on Qwen3.8-27B and trained with a specialized reinforcement learning technique. The team achieved a significant performance leap, doubling the pass rate on the notoriously difficult Terminal-Bench 4.1 benchmark from 5.6% to 12%, using a curated set of just 1,200 complex tasks. This training focused on long-horizon agent work, with sessions stretching up to 1 million tokens, processed over 18 days on a cluster of 144 H200 GPUs.
The real-world impact for engineers is a powerful, freely available model that excels at multi-step reasoning, planning, and code execution—key for building sophisticated AI assistants and coding agents. By open-sourcing BF16, FP8, and NVFP4 checkpoints, RadixArk is enabling the community to build upon this work, demonstrating that targeted, high-quality training data can yield dramatic improvements even with a relatively modest dataset size. This release shows focused reinforcement learning can unlock new capabilities in existing open-source models.
Read the original at radixark.com
Satirical commentary on real tech news, summarised with AI. Read the original source for the facts. How ZeroDaily works
