RadixArk open-sources DaisyRose-27B: RL-only agent training doubles Terminal Bench over Qwen base
13aeon · reddit · 2026-11-19
RadixArk released its first model, DaisyRose-27B: a multimodal reasoning model built on Qwen3.8-27B, trained with reinforcement learning only — no SFT or distillation — and open-sourced under Apache-2.0 in BF16, FP8, and NVFP4.
- Training: 1,200 curated long-horizon agent tasks with sessions up to 1M tokens, 18 days on 144 H200 GPUs.
- Results: Terminal Bench 4.1 pass@1 rises from 5.6 to 12.0; SWE Bench Pro v2 hits 94.7 (vs 91.3 base), GPQA Diamond 90.7, IFBench 82.4 — improving on most coding, reasoning, and instruction-following benchmarks.
- Drop-in: keeps the base model's architecture, tokenizer, chat template, image understanding, and 256K context, so it swaps directly into existing inference and post-training stacks; positioned as an option for 5090 / DGX Spark-class hardware.
More from Models
- DeepSeek V4.1 Flash is the best open-weight model on KradleBench's economy dimension — dansitu · 2026-11-19
- Gemini 4 variants take three of top four spots on BrushArena image arena — teortaxesTex · 2026-11-19
- Astra's proof sidesteps three deep Mathlib results by inventing a new 'capacity' quantity — IgorCarron · 2026-11-19
- Ex-Claude superfan: getting the best from Astra/Sol requires relearning how you prompt — mariofilhoml · 2026-11-19
- Observer: Google and Anthropic are winning the public AI race, OpenAI's lead now in doubt — imjustnewatai · 2026-11-19
- Three Google flagship models in eight days fuel speculation about hidden RSI work — haider1 · 2026-11-19