RadixArk's DaisyRose-27B lifts Terminal-Bench 4.1 from 5.6% to 12.0% with 18-day RL run on 144 H200s
BanghuaZ · x · 2026-11-19
RadixArk released DaisyRose-27B, its first model built on Qwen3.8-27B and trained with RL using its Miles framework.
- Training: only 1,200 curated agent tasks built with Eigent AI, sessions up to 1M tokens, 18 days on 144 H200 GPUs.
- Results: Terminal-Bench 4.1 pass@1 improved from 5.6% to 12.0%, approaching GPT-6 Luna's 15.2%, with gains across reasoning, instruction-following and coding benchmarks.
- Efficiency: 37% MFU on 27B dense, 11.7% on 300B MoE, and 80% lower sandbox costs.
- Task design: each task targets a known mistake of the base model, with difficulty tuned to a suitable solve-rate range.
- Open release: BF16, FP8 and NVFP4 checkpoints, runnable on an RTX 5090 or DGX Spark.
Related event: RadixArk Open-Sources DaisyRose-27B, Doubling Terminal-Bench Scores(6 posts)→
More from coding & agent
- CodeRabbit launches Verify beta: agentic testing and AI QA for agent-written code — rohanpaul_ai · 2026-11-19
- StreamEmbed: On-device real-time visual indexing of your screen history, 3.7x faster — CShorten30 · 2026-11-19
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Pierre releases Words, a streaming Markdown parser 2x faster than md4x and 20x faster than Streamdown — andersonbcdefg · 2026-11-19
- Ex-Claude superfan: getting the best from Astra/Sol requires relearning how you prompt — mariofilhoml · 2026-11-19
- Rivet's lesson: JEV doesn't reason—split severity classification into tiny yes/no prompts — yawnxyz · 2026-11-19