{"slug": "daisyrose-27b-pushing-the-frontier-of-post-training-efficiency", "title": "DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency", "summary": "RadixArk's DaisyRose-27B doubled Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% using 1,200 curated tasks and 18 days of training on 144 H200 GPUs, according to the model's Hugging Face release post. The Apache-2.0 checkpoint retains Qwen3.8-27B's 256K context window, adds default chain-of-thought reasoning, and is positioned as a drop-in upgrade that runs with SGLang plus DSpark speculative decoding for roughly 2x lower latency than vanilla vLLM. The release indicates targeted reinforcement learning on a small, high-quality dataset can lift long-horizon agent performance without scaling data or compute.", "body_md": "[Hugging Face](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin)\n\n### DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency\n\nWhich summary reads better? Pick one — models revealed after.Both summaries are AI-generated.\n\nDaisyRose-27B doubles Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% using just 1,200 curated tasks and 144 H200s for 18 days. This means you can now ship agentic workflows with 2x higher success rates on long-horizon tasks without scaling data or compute—just drop in the Apache-2.0 checkpoint and serve it with SGLang + DSpark speculative decoding for ~2x lower latency than vanilla vLLM.\n\nDaisyRose-27B doubles Terminal-Bench 4.1 pass@1 performance from 5.6% to 12.0% with just 1,200 curated tasks and 18 days of training on 144 H200 GPUs. This shows how targeted reinforcement learning on small, high-quality datasets can dramatically improve agent capabilities without massive compute, enabling faster iteration on specialized agent workflows in production. The model retains Qwen3.8-27B's 256K context window and adds default chain-of-thought reasoning, making it a drop-in upgrade for existing deployments needing better long-horizon task performance.", "url": "https://wpnews.pro/news/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency", "canonical_source": "https://www.snipvote.com/story/cmwn879xx0004b5bg3dc7w8qj", "published_at": "2026-11-19 07:43:08.368313+00:00", "updated_at": "2026-11-19 07:43:10.707598+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-agents", "ai-research"], "entities": ["RadixArk", "DaisyRose-27B", "Qwen3.8-27B", "Terminal-Bench 4.1", "Hugging Face", "SGLang", "DSpark", "vLLM"], "also_reported_by": [], "takeaway": "Reinforcement learning on 1,200 curated tasks over 18 days on 144 H200s doubled Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%, and serving the Apache-2.0 checkpoint with SGLang plus DSpark speculative decoding cut latency to about half of vanilla vLLM.", "recommendation": {"advice": "When post-training an agent model, invest in a small curated task set rather than scale: 1,200 tasks and 18 days on 144 H200s doubled Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.", "playbook": "training-data", "playbook_url": "https://wpnews.pro/playbooks/training-data"}, "alternates": {"html": "https://wpnews.pro/news/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency", "markdown": "https://wpnews.pro/news/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency.md", "text": "https://wpnews.pro/news/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency.txt", "jsonld": "https://wpnews.pro/news/daisyrose-27b-pushing-the-frontier-of-post-training-efficiency.jsonld"}}