DaisyRose-27B: RL on 144 H200s for 18 days lifts Terminal-Bench from 5.6% to 12.0%
ying11231 · x · 2026-11-19
RadixArk released its first open-weight model, DaisyRose-27B: RL trained on Qwen3.8-27B with 1,200 curated long-horizon agent tasks across 18 days on 144 H200s, stretching agent sessions to 1M tokens.
Key results:
- Terminal-Bench 4.1 score raised from 5.6% to 12.0%, approaching GPT-6 Luna's 15.2%, with gains across all other benchmarks evaluated
- Infra: 37% MFU on 27B dense, 11.7% on 300B MoE, 80% lower sandbox costs
- The 1.2K task suite was built with EigentAI
- Open weights, runnable on a single RTX 5090 or DGX Spark
A concrete proof that small models + curated RL tasks can close much of the gap to larger models, with the training recipe and engineering data shared publicly.
More from Models
- Google engineer says Gemini Neon edges out Opus 5.5 Max on Agents Arena — lennysan · 2026-11-19
- Gemini 4 Neon appears live and quick: researcher shares one-shot demos, best scores 7/10 — SamuelAlbanie · 2026-11-19
- Astra 6.2 Day Two: Real-World Usage Far Worse Than Benchmarks, But Cheaper Than 6.1 — bindureddy · 2026-11-19
- Early Testers Pan GPT-6.2 Astra — and Praise OpenAI for Not Silencing Them — paw_lean · 2026-11-19
- Hands-on Gemini 4 Neon: SDF dragon, walkable house from a 1959 floor plan, no-training neural nets — Prompt Engineering · 2026-11-19
- Gemini 4 Neon costs $6.44 per task, 17% below Carbon but 1.6x GPT-6.2 Astra — ArtificialAnlys · 2026-11-19