DaisyRose-27B RL boosts Qwen3.8 on Terminal-Bench
x.com

DaisyRose-27B RL boosts Qwen3.8 on Terminal-Bench

The gist

RadixArk released DaisyRose-27B, an open-weight model that lifted Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.
The 18-day, 144-GPU run tests whether small curated RL tasks generalize or just tune the benchmark.

How it unfolded

  1. Nov 18, 2026RadixArk launched DaisyRose-27B, built on Qwen3.8-27B and trained with reinforcement learning using its open-source Miles framework.
  2. Nov 18, 2026Training ran 18 days on 144 H200 GPUs over about 1,200 curated long-horizon agent tasks, with sessions up to 1M tokens, consuming 63,000 GPU-hours.
  3. Nov 18, 2026Terminal-Bench 4.1 pass@1 rose from 5.6% to 12.0%, a 114% relative improvement, with the largest gain on that benchmark.

Key terms

open-weight
Numerical values in a neural network that determine how strongly inputs contribute to its calculations.
Terminal-Bench
A benchmark test suite used to measure how well AI models perform on terminal-based coding and command-line tasks.
pass@1
A clustering method that partitions observations into k groups, each assigned to the nearest cluster mean.

Source: Wikipedia

Sources

Read the full story in the app Gistline — free on the App Store