
Apache 2.0 Post-Training Release Built on Qwen3.8-27B
RadixArk has released DaisyRose-27B, its first open-weight model, applying long-horizon reinforcement learning to Alibaba's Qwen3.8-27B base architecture. Detailed in a Hugging Face technical announcement published on November 18, 2026, the release targets multi-step software engineering, terminal automation, and scientific reasoning tasks. RadixArk published three Apache 2.0 checkpoints on Hugging Face—standard BF16, FP8, and NVFP4—giving self-hosted developers quantized and full-precision options without restrictive commercial terms.
Rather than altering the underlying foundation model's structure, DaisyRose-27B retains Qwen3.8-27B's architecture, tokenizer, chat template, image understanding capabilities, and 256K-token context window. While the model's native window remains 256K tokens, RadixArk trained it across agent sessions reaching up to 1 million tokens by using harness summarization to fold extended interactions into multiple trajectories. According to RadixArk, the entire post-training run relied on roughly 1,200 curated tasks executed over 18 days on a 144-GPU NVIDIA H200 cluster, totaling 63,000 GPU-hours.
Asynchronous RL Architecture, AgentENV Sandboxes, and Benchmarks
To sustain long-horizon agent rollouts without stalling gradient updates, RadixArk ran its open-source reinforcement learning framework, Miles, in fully asynchronous mode. Across the 144 H200 GPUs, 128 accelerators continuously executed agent inference tasks while 16 GPUs ran the trainer and broadcast updated weights back to the inference pool after every optimization step. Instead of relying solely on stock open-source trainers, RadixArk paired Miles with an internal training backend engineered for long-context agent trajectories.
In its release report, RadixArk claims this internal backend achieved 37% model FLOPs utilization (MFU) during the DaisyRose-27B run, compared with 16% MFU on a Megatron-based open-source trainer executing comparable workloads—a 57% reduction in GPU time. End-to-end throughput, factoring in the time agents spent executing tools, reportedly reached roughly 16,800 tokens per second across 144 GPUs versus about 8,300 tokens per second for the open-source baseline. RadixArk also reports larger MFU gains on mixture-of-experts architectures tested internally, citing 6.2x improvements on Qwen3.6-35B-A3B and over 10x on the roughly 300B-parameter GLM-5.3-Flash.
Training data came from a partnership with Eigent AI, narrowing thousands of candidate prompts across machine learning, software engineering, and science down to about 1,200 tasks. Each task was calibrated around failures observed in Qwen3.8-27B, allowing up to 1.5 hours in a sandbox with more than 100 tool calls and over 256K output tokens, and audited to eliminate shortcut passes. Environment execution ran on an open-source AgentENV sandbox service scaled from one to 12 machines with dynamic per-sandbox memory allocation. Between mid-September and mid-November, the service launched roughly 944,000 sandboxes—peaking at 1,566 concurrent instances—with a 99% overall creation success rate (above 99.95% since late October) and an estimated 80% cost reduction versus commercial cloud sandbox providers.
On RadixArk's evaluation suite, DaisyRose-27B outperformed Qwen3.8-27B across seven reported benchmarks. The largest relative jump appeared on Terminal Bench 4.1, where pass@1 rose from 5.6% to 12.0%. Other reported gains include SWE Bench Pro v2 (91.3 to 94.7), DeepSWE v1.1 (42.2 to 46.2), IFBench (79.5 to 82.4), Terminal Bench 2.1 (73.0 to 75.5), GPQA Diamond (89.2 to 90.7), and text-only HLE (33.9 to 34.3).
Developer Serving Workflows with SGLang and DSpark
For local and cluster deployment, RadixArk recommends serving DaisyRose-27B using SGLang paired with DSpark speculative decoding, backed by launch configurations in the SGLang cookbook for varying GPU topologies and quantization formats. Because the server exposes an OpenAI-compatible HTTP endpoint—defaulting to port 30000 in the project's quickstart—teams can integrate the model into existing agent harnesses and IDE pipelines through standard OpenAI Python client libraries without rewriting orchestration code.
By default, DaisyRose-27B generates an explicit reasoning trace before emitting its final response. Through the chat completions API, developers receive the step-by-step trace in the reasoning_content field and the actionable output in content. RadixArk's recommended baseline sampling parameters specify a temperature of 1.0, top_p of 0.95, top_k of 20, and min_p of 0.0, aligning with the model card's inference profile for complex tool-use and coding workflows.
Reproducibility Caveats, Quantization Notes, and Licensing
While Apache 2.0 weights and open-source components such as Miles, SGLang, AgentENV, slime, and Harbor lower barriers for adoption, open-source practitioners face notable reproducibility and evaluation caveats. Most importantly, the 37% MFU and 2x throughput figures depend on RadixArk's proprietary internal training backend rather than the public Megatron-based open-source trainer, meaning external teams attempting to replicate the 18-day H200 training budget with fully open tooling should expect lower utilization.
In addition, RadixArk's benchmark table reports the higher score between BF16 and FP8 evaluations for each metric rather than publishing separate full tables for all three checkpoints, though the team notes that BF16 and FP8 differed by less than one point wherever both were run. Performance of the smaller NVFP4 checkpoint (listed at 18B parameters on Hugging Face versus 28B for BF16 and FP8) is not broken out in the headline comparison, and HLE results reflect text-only evaluation despite the model retaining multimodal image-text-to-text capabilities. Independent third-party verification will be needed to confirm real-world reliability across uncurated repositories and non-sandboxed developer environments.
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.



