Hands-on with DaisyRose-27B: pure-RL fine-tune edges out Qwen3.8-27B
DerTomsn · reddit · 2026-11-21
- RadixArk released DaisyRose-27B, a Qwen3.8-27B fine-tune trained with reinforcement learning only — no SFT, no distillation — on 1,200 long agent tasks. It keeps the base architecture, tokenizer, chat template, 256K context and vision encoder; Apache-2.0.
- The author benchmarked it against the base model on identical hardware/settings (oQ8e-mtp quant, xhigh reasoning): code 81.3 vs 81.0, agent workflow 89.3 vs 88.2, roleplay 93.9 vs 92.1, research 87.8 vs 86.4 — overall 88.1 vs 86.9. A consistent 1-point win, within run-to-run spread.
- Speed (50 tok/s) and memory (38 GB) are unchanged; the cost is 5% more reasoning tokens per run.
- Vendor's claimed Terminal Bench 4.1 pass@1 doubling (5.6→12.0) targets long-horizon agent work that llm-bench.io doesn't test, so it remains unverified; real-world agentic gains may exceed the 1-point quality delta.
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22