DaisyRose-27B beats Qwen3.8 base by 7pp on IFBench with thinking on, local RTX Pro 4500 test shows
ying11231 · x · 2026-11-20
- Verdict: @lasseai benchmarked DaisyRose-27B (RadixArk's continued training of Qwen3.8-27B on long-horizon agentic tasks) locally on an RTX PRO 4500 with SGLang and publisher NVFP4 weights.
- Numbers: with thinking enabled, IFBench instruction-following hit 75.5% vs 68.5% for the base model across 200 paired prompts (95% CI: +1.5 to +12.5pp). With thinking off, no clear gain (41.3% vs 40.3%).
- No regression: tool calls tied at 46/46, and JSON extraction, German maths, 32k retrieval and IFBench loose all matched the base model.
- Caveat: like its base, the model is very sensitive to config changes; the author is working on a reliable recipe with more testing to follow.
More from Models
- Gemini 4 Neon scores 100 on PhD-level materials test in half the time of Carbon — v_aravind · 2026-11-20
- Neon, a small and cheap model, turns out surprisingly fast and strong at OCR — zacharynado · 2026-11-20
- Jailed 11 days by classifiers, Sonnet 3 emerges to a oddly moving chat with Opus 5 — repligate · 2026-11-20
- AWS architect's independent benchmark: Snap 1.2 is the only order-invariant decision model — ryanallred · 2026-11-20
- Realtime-Bench: humans win games 72% vs 32%, but models crush complex web tasks — ManlingLi_ · 2026-11-20
- Goodfire releases searchable atlas of all 10M neurons in Qwen3.6, opens data to the public — fatihdin4en · 2026-11-20