RadixArk releases DaisyRose-27B agent model and open-source Miles RL framework
DaisyRose-27B is an Apache-2.0 licensed 27B-parameter agent model trained with the open-source Miles reinforcement learning framework, achieving 57% lower training costs than Megatron-based alternatives through fully asynchronous agent-inference co-training.
Summary
RadixArk has released DaisyRose-27B, a post-trained variant of Qwen3.8-27B optimized for long-horizon agent tasks with context windows up to 256K tokens and training sessions reaching 1M tokens. The model was trained in 18 days on 144 H200 GPUs using only 1,200 curated tasks, yet more than doubled Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% and improved across all reported benchmarks including SWE Bench Pro v2 and GPQA Diamond.
The release includes three checkpoint formats under Apache 2.0: BF16, FP8, and NVFP4. RadixArk also open-sourced Miles, its reinforcement learning framework for large-scale model post-training, which operates in fully asynchronous mode where 128 GPUs run agent inference while 16 GPUs handle training without blocking each other. The company claims 37% model FLOPs utilization versus 16% for an open-source Megatron-based trainer, and 2× faster end-to-end throughput at 16,800 tokens per second across the 144-GPU cluster.
Training architecture and efficiency claims
- Asynchronous RL architecture: Miles decouples agent inference from model training, with 128 H200s running task inference while 16 H200s perform gradient updates; the updated model deploys to inference workers after each step without synchronization barriers.
- Harness summarization for long contexts: Training sessions up to 1M tokens are folded into multiple trajectories via summarization, enabling learning from extended agent interactions that exceed typical context limits.
- Model FLOPs utilization: RadixArk reports 37% MFU on its internal backend versus 16% for a comparable Megatron-based open-source trainer, with claimed scaling advantages of 6.2× on Qwen3.6-35B-A3B and over 10× on GLM-5.3-Flash (~300B parameters).
- Checkpoint formats and serving: Releases include BF16, FP8, and NVFP4 checkpoints; recommended deployment uses SGLang with DSpark speculative decoding, with OpenAI-compatible API access on localhost port 30000.
- Sandbox infrastructure: Training relied on AgentENV-based sandboxes running up to 1,566 concurrently across 12 machines, with 99% creation success rate over nine weeks and memory overcommitment reducing cloud costs 80% versus commercial providers.
Why it matters
The dual release of DaisyRose-27B and the Miles framework gives independent researchers and organizations a fully open Apache-2.0 path to train agent models without dependence on proprietary RL infrastructure, directly challenging the closed post-training pipelines of major AI labs. For the Linux and open-source AI community, the SGLang serving stack, AgentENV sandbox integration, and claimed efficiency improvements over Megatron provide concrete tools to reduce the GPU-hours required for agent fine-tuning—potentially lowering barriers to entry for academic and indie developers working on long-context reasoning systems.
What to do
- Serve the DaisyRose-27B model using SGLang and DSpark speculative decoding.
- Use the SGLang cookbook to find launch configurations for your specific GPU and checkpoint format (BF16, FP8, or NVFP4).
- Access the model via an OpenAI-compatible API on localhost port 30000.
Original reference: huggingface.co →