RadixArk released DaisyRose-27B, an open-weight agentic coding model post-trained from Qwen3.8-27B using reinforcement learning (RL) exclusively, with no supervised fine-tuning or distillation. The 27.8B-parameter dense architecture retains Qwen3.8-27B's structure: 64 decoder layers in a 3:1 pattern alternating three Gated DeltaNet linear-attention layers with one full-attention layer, a 27-layer vision encoder (0.46B params), and a multi-token prediction head (0.42B params) for speculative decoding, supporting 262,144-token context. Training leveraged ~1,200 curated tasks from Eigent AI, each allowing up to 1.5 hours, 100+ tool calls, and 256k output tokens within AgentENV sandboxes (944,000 runs, 99% creation success, peak 1,566 concurrency). The 18-day run used 144 H200 GPUs (128 inference, 16 training) totaling 63,000 GPU-hours on RadixArk's Miles framework, achieving 37% MFU and a 2.0x end-to-end throughput speedup (16.8k vs 8.3k tokens/s) over a Megatron baseline. Benchmarks show Terminal Bench 4.1 pass@1 more than doubling from 5.6 to 12.0, with steady coding gains: DeepSWE v1.1 +4.0 (46.2), SWE Bench Pro v2 +3.4 (94.7), and IFBench +2.9 (82.4, top of RadixArk's 9-model table). Knowledge gains are marginal: HLE text-only +0.4 (34.3) and GPQA Diamond +1.5 (90.7). DaisyRose-27B trails larger MoE models like MiMo V2.6 Flash (309B) on Terminal Bench 4.1 (16.1) and Inkling 2 Small (594B) on DeepSWE (72.1) and HLE (44.1). Deployment supports BF16 (~52 GB), FP8 (~29 GB), NVFP4 (~23 GB, Blackwell only), and GGUF on a single GPU; RadixArk recommends SGLang with DSpark speculative decoding (temperature 1.0, top_p 0.95, top_k 20). All scores are self-reported under Apache 2.0.
Read original
marktechpost