DaisyRose-27B
A new open-weight model called DaisyRose-27B, a heavily post-trained variant of Qwen 3.8 27B, is drawing attention for its strong benchmark scores and substantial training run (18 days on 144 H200 GPUs), but some users question whether the gains reflect real capability or “benchmaxxing” tuned to specific evals. Commenters dig into hardware cost estimates, quantization and deployment on consumer GPUs, and whether targeted RL-style post-training can genuinely improve models across the board without hidden regressions in areas like prose, metaphor understanding, or generality. The thread situates DaisyRose within a broader trend of third parties post-training open models, while also surfacing concerns about eval-data dependence, data privacy, and the limits of current approaches to AGI and self-improvement.
Performance and “benchmaxxing” concerns
- Some users test DaisyRose-27B against private eval sets.
- One report: DaisyRose slightly outperforms another Qwen 3.8-based finetune on an 86-task “medium reasoning” bench (0.898 vs 0.87) but uses more VRAM, longer context, different quantization, and more tokens.
- Given DaisyRose’s advantages (KV cache, context, quantization, decode speed, hardware), this gain is viewed by some as marginal and labeled “benchmaxxing.”
- Others note the blog shows multiple benchmarks improving without regressions, but skeptics still expect hidden tradeoffs in unmeasured areas (e.g., prose, metaphor understanding, prompt interpretation).
Training setup, topology, and cost
- Training reportedly used 144 H200 GPUs for 18 days (~63k GPU-hours).
- Back-of-envelope price estimates on current markets put cost around mid–hundreds of thousands of dollars, with wide variation by provider.
- Discussion on whether cheaper accelerators (e.g., other vendors’ chips) would help is contentious; some argue if there were a clearly cheaper, reliable path, it would already be used.
- Authors state they used dedicated rollout and trainer nodes with RDMA between them.
Quantization, formats, and consumer hardware
- Initial frustration about lack of GGUF; a community GGUF conversion is quickly shared.
- Clarification that the model is already quantized but needs conversion to GGUF for llama.cpp-style use.
- Qwen 27B–based models are praised for “punching above their weight” and fitting on higher-end consumer machines (e.g., Mac Mini / laptop with enough RAM or 48GB VRAM GPUs).
Post-training strategy and generalization
- DaisyRose is a heavy post-training/RL run on top of Qwen 3.8 27B, itself already post-trained from 3.6.
- The process is described as gap-finding plus high-quality targeted data, especially to boost Terminal Bench 4.1.
- Some argue focused RL can improve targeted tasks and still generalize modestly to related benchmarks.
- Others believe heavy post-training inevitably narrows models, making them better at coding/eval tasks but worse at messy, unbenchmarkable language tasks.
Broader ecosystem and future directions
- Comparisons are made to other post-trained open models (e.g., Ember, Particle, Nex, Sabi-noema) with mixed results.
- There’s demand for more strong, open-weight base models beyond Qwen and for variants optimized for specific languages or backend tasks.
- Some call for a model that aggregates improvements from multiple post-training efforts, rather than many slightly-better-than-Qwen forks.