RadixArk trains DaisyRose-27B in 18 days on 144 H200s, doubling Terminal-Bench score
BanghuaZ · x · 2026-11-19
RadixArk released its first open-weight model, DaisyRose-27B: RL training on top of Qwen3.8-27B with 1,200 curated long-horizon agent tasks, using 144 H200s for just 18 days.
- Terminal-Bench 4.1 score lifted from 5.6% to 12.0% (the shared chart also compares Muse Glimmer-30B, GPT-6 Luna and others)
- Agent sessions stretched up to 1M tokens of context
- Built on the open-source Miles framework plus their own training backend
A reference case of a small team massively boosting a small model's agent capability with short-cycle RL.
More from coding & agent
- 130,000+ SEC Form 144 filings added to a dataset, revealing insider sales up to 4 days before Form 4 — virattt · 2026-11-19
- Six deployment strategies compared: rolling, blue-green, canary and more — JeremyCMorgan · 2026-11-19
- New MAC Bench finds frontier agents coordinate poorly: 8-agent teams win by exploration, not teamwork — AlexGDimakis · 2026-11-19
- Dev Argues AI Needs One Interface: Claude or GPT Plus Filesystem or Sandbox — _AustinCalvert_ · 2026-11-19
- 19 Core Concepts of AI System Design: From KV Cache to Multi-Agent Systems — blaizedsouza · 2026-11-19
- 10 Free Repos That Teach AI Agent Harness Engineering: Tools, Memory, Rules — blaizedsouza · 2026-11-19