DaisyRose-27B: 1,200-task RL lifts Qwen3.8 Terminal-Bench 4.1 from 5.6% to 12.0%Image: radixark.com
Cooling

DaisyRose-27B: 1,200-task RL lifts Qwen3.8 Terminal-Bench 4.1 from 5.6% to 12.0%

Heat 46 · #33 in AI · Cooling2 reports · 2 sources · First reported Nov 18 · Updated Nov 20

The story

RadixArk has released DaisyRose-27B, its first model, built by reinforcement-training Qwen3.8-27B on 1,200 curated long-horizon agent tasks. According to RadixArk, training used its Miles RL framework, ran 18 days on 144 H200 GPUs, and allowed sessions of up to one million tokens. The company reports the model's pass@1 score on Terminal-Bench 4.1 rose from 5.6 percent to 12.0 percent, with further gains on reasoning, instruction-following and coding benchmarks; the figures are vendor self-reported. Checkpoints are released under Apache 2.0 in BF16, FP8 and NVFP4 formats.
Latest development
record: round 213 log text '66.0->62.0' was a typo, script set (60,80,50)=66.0 and CSV never changed

How it unfolded

  1. record: round 213 log text '66.0->62.0' was a typo, script set (60,80,50)=66.0 and CSV never changed
  2. benchmark figures are vendor self-reported
  3. Hugging Face blog carries the same announcement (second platform)
  4. Terminal-Bench 4.1 pass@1 5.6% -> 12.0%, gains on reasoning, instruction-following and coding benchmarks
  5. Nov 18
    RadixArk official blog 2026-11-18: first model from RadixArk, built on Qwen3.8-27B, trained with RL framework Miles on 1,200 curated long-horizon agent tasks (sessions up to 1M tokens), 18 days on 144 H200 GPUs

The angle

1,200 tasks, 18 days, 144 GPUs: how a small RL run doubled a 27B model's Terminal-Bench score, and what the vendor numbers don't proveSuggested format: Short video · for AI builders, open-model watchers

Heat

Now 46 · Peak 46 · 64 recorded checks since Nov 20
Full ranking →