What RadixArk released

RadixArk has published DaisyRose-27B, its first open-weight model, described as an agentic coding model derived from Qwen3.8-27B. According to the model card reported by the source, post-training is based exclusively on reinforcement learning, with no supervised fine-tuning or distillation, and it is distributed under the Apache 2.0 license as a direct replacement for the base model.

The architecture remains that of the base model: 64 decoder layers with a 3:1 pattern, in which three Gated DeltaNet linear-attention layers alternate with one full-attention layer, hidden size 5,120, a 27-layer vision encoder and a multi-token prediction layer for speculative decoding. Input is text and images, output is text, and the model is described as a thinking model. The stated context window is 262,144 tokens.

How it was trained

Training took place on the Miles platform, RadixArk's open-source RL framework, in fully asynchronous mode. The source reports 144 H200 GPUs used for 18 days, of which 128 were dedicated to agent inference and 16 to training, for a total of 63,000 GPU-hours. RadixArk also claims an MFU of 37% on its own backend, versus 16% for an open-source Megatron-based trainer, and end-to-end throughput of 16,800 tokens per second versus 8,300, with a 57% reduction in GPU time.

The training corpus comprises approximately 1,200 tasks curated with Eigent AI, selected from thousands of candidates and designed to target specific weaknesses of the base model. Each sandbox attempt allows up to 1.5 hours, over 100 tool calls and more than 256,000 output tokens; agent sessions reach up to 1M tokens. The sandboxes ran on AgentENV: RadixArk reports about 944,000 environments created between mid-September and mid-November, peak concurrency of 1,566 and a 99% successful-creation rate over nine weeks, with a claimed cost 80% lower than the most widely used cloud providers.

Reported results and limitations

All scores are self-reported by RadixArk and, for DaisyRose, correspond to the best value among the BF16 and FP8 runs. The most marked gain is on Terminal Bench 4.1, where pass@1 rises from 5.6 to 12.0; next are DeepSWE v1.1 from 42.2 to 46.2, SWE Bench Pro v2 from 91.3 to 94.7 and IFBench from 79.5 to 82.4. Improvements in knowledge are more modest: HLE, text-only, rises from 33.9 to 34.3 and GPQA Diamond from 89.2 to 90.7.

The comparison with larger models is mixed. MiMo V2.6 Flash, with 309 billion total parameters, leads Terminal Bench 4.1 with 16.1; Inkling 2 Small, with 594 billion, leads DeepSWE with 72.1 and HLE with 44.1. In the table of nine models published by RadixArk, DaisyRose-27B ranks first only on IFBench. The weights are available on Hugging Face in BF16, FP8, NVFP4 and GGUF; the NVFP4 variant requires a Blackwell GPU. RadixArk recommends SGLang with DSpark speculative decoding, based on the Qwen3.8-27B-DSpark draft model, and notes compatibility with vLLM as well. The claimed performance and comparison with competing models remain to be independently verified.

Source and attribution: this original summary was prepared by AICronaca based on the report by MarkTechPost ↗. The statements are attributed to the source; any informational limitations: the reconstruction follows the information available in the original report.