Start typing to search this publication.
Tech Workflow Hub
Open menu

Subscribe to Tech Workflow Hub

Get new posts delivered straight to your inbox.

RadixArk DaisyRose-27B Explained: How Pure RL (No SFT) Doubled Terminal-Bench 4.1 Agent Scores

Inside the open-weight model post-trained purely through reinforcement learning on long-horizon agent tasks.

Tech Workflow Hub
1 min read

Released November 18, 2026 under the Apache 2.0 license, DaisyRose-27B is RadixArk's first flagship open-weight model. What sets DaisyRose-27B apart from conventional frontier models is its training paradigm: it was post-trained strictly via Reinforcement Learning (RL only), entirely bypassing traditional Supervised Fine-Tuning (SFT) and distillation pipelines.

Architecture & Training Methodology

  • Base Model: Built directly on Qwen3.8-27B (dense 27B architecture with native vision capabilities and a native 256K token context window).

Agentic RL Training: Trained on approximately 1,200 curated, complex, long-horizon agent tasks with multi-step sessions scaling up to 1M tokens.

Compute Infrastructure: Trained over 18 days on a cluster of 144 NVIDIA H200 GPUs using RadixArk's open-source Miles framework.

Benchmark Results & Production Serving

In real-world coding and environment execution evaluations, DaisyRose-27B demonstrated a massive performance leap. On Terminal-Bench 4.1, it boosted pass@1 accuracy from 5.6% (the base model score) to 12.0%, proving that pure RL without SFT can achieve state-of-the-art multi-step reasoning.

For inference and enterprise deployment, RadixArk released weights in BF16, FP8, and NVFP4 formats. Production serving is recommended via SGLang paired with DSpark speculative decoding for low latency under high concurrency.

Agentic Efficiency in 2026

As frontier development shifts from massive pre-training to post-training inference optimization, agentic reasoning density has become the key differentiator. To see how architectural efficiency impacts real-world token budgets and deployment costs, read our Qwen4-Max 1M context efficiency analysis.

Subscribe to Tech Workflow Hub

Practical AI workflows, software benchmarks, and automation guides.