DaisyRose-27B Doubles Qwen3.8 Terminal-Bench Score

Officially launched on November 18, 2026.
RadixArk officially launched DaisyRose-27B on November 18, 2026, releasing the first publicly available model from the company's coding-agent research program. Built on Qwen3.8-27B and trained with reinforcement learning using the open-source Miles framework, DaisyRose-27B more than doubles the base model's Terminal-Bench 4.1 pass@1 score, climbing from 5.6% to 12.0%. The release includes BF16, FP8, and NVFP4 checkpoints on Hugging Face, offering researchers and developers a concrete data point for coding-agent RL recipes on open-weight foundations.
RadixArk trained DaisyRose-27B on 1,200 curated long-horizon agent tasks over 18 days using 144 H200 GPUs. Sessions extended up to 1 million tokens, designed to test multi-step reasoning and tool use across machine learning, software engineering, and scientific domains. The company partnered with Eigent AI to build the training dataset, selecting tasks based on three criteria: difficulty gradient, realism in real-world coding workflows, and coverage of diverse skill requirements. Each task ran in isolated sandboxes to grade the agent's attempts safely and reliably.
Training Infrastructure and Efficiency Gains
The training backend achieved 37% model FLOPs utilization (MFU) on this release, representing a 57% reduction in GPU time compared to 16% MFU on comparable open-source trainers built on Megatron. RadixArk's fully asynchronous architecture allows the agent to continue working while the model learns, with 128 H200s running inference on agent tasks and 16 running the trainer. After every training step, the updated model deploys immediately to the inference fleet. End-to-end throughput reached 16,800 tokens per second across all 144 GPUs, twice the speed of reference implementations on identical hardware.
The training run consumed 63,000 GPU-hours total, maintaining stable speed and efficiency from the first step to the last. RadixArk's backend handles mixture-of-experts models with even larger efficiency multipliers: 6.2× on Qwen3.6-35B-A3B and more than 10× on the roughly 300-billion-parameter GLM-5.3-Flash. The company designed the system for agent training specifically, learning from a small number of very long sessions folded into multiple trajectories via harness summarization rather than batching many short sequences.
Benchmark Performance Across Tasks
DaisyRose-27B outperformed the base Qwen3.8-27B model on every benchmark RadixArk tested. Beyond the headline Terminal-Bench 4.1 result, the model showed gains on Terminal-Bench 2.1, DeepSWE v1.1, SWE Bench Pro v2, HLE text-only tasks, IFBench instruction following, and GPQA Diamond scientific reasoning. Terminal-Bench 4.1 remains a notoriously difficult long-horizon benchmark, and the 114% relative improvement demonstrates the impact of focused RL training on extended coding workflows.
RadixArk evaluated Terminal-Bench 4.1 by averaging 8 runs per task using an internal mini-swe-agent harness with summarization enabled. Terminal-Bench 2.1 used 5 runs per task with the Terminus 2 harness. DeepSWE v1.1 and SWE Bench Pro v2 averaged 8 and 5 runs per task respectively, both using the mini-swe-agent harness. HLE text-only tasks ran 5 times per problem, graded by GPT-6 Luna without tools. IFBench averaged 4 runs on each of 300 prompts, scored by the official allenai/IFBench scorer for prompt-level loose accuracy. GPQA Diamond averaged 8 runs on each of 198 questions using the simple-evals multiple-choice template with reshuffled answer order.
Sandbox Infrastructure at Scale
Every agent attempt during training ran in a fresh sandbox copy of the task environment, isolating command execution, file edits, and grading. RadixArk's sandbox service, built on the open-source AgentENV project, executed approximately 944,000 sandboxes for this release between mid-September and mid-November 2026, with peaks of 1,566 concurrent sandboxes. The service scaled from one machine to 12 during the training period, shipping ten upgrades with minimal downtime. By late October, sandbox creation success rates exceeded 99.95%.
The team kept cloud infrastructure costs for sandboxes roughly 80% lower than popular cloud sandbox providers by allocating memory dynamically. Each sandbox receives memory that grows and shrinks with actual usage rather than static allocation, allowing each machine to run significantly more sandboxes concurrently. RadixArk delivered steady reliability improvements throughout the training run, including fixes to prevent machines from freezing when memory runs out.
Availability and Open-Source Components
DaisyRose-27B checkpoints are available immediately on Hugging Face in BF16, FP8, and NVFP4 formats. The Miles reinforcement learning framework used to train the model remains open source, enabling researchers to replicate and extend the training approach. RadixArk emphasized that the release demonstrates what high-quality data and efficient infrastructure can achieve on limited compute budgets, positioning the work as a template for teams building specialized models. The company is now working with teams exploring post-training for domain-specific applications, focusing on improving targeted capabilities while preserving broader model strengths.
RadixArk credited Eigent AI for training data collaboration and the Qwen team for creating the Qwen3.8-27B foundation model. The company also acknowledged contributions from the Miles and SGLang communities, along with open-source projects including slime, AgentENV, and Harbor.
What This Means
DaisyRose-27B establishes a replicable path for small teams to achieve meaningful gains on coding-agent benchmarks without frontier-scale compute. The 57% training cost reduction and 2× end-to-end speed improvement over reference implementations lower the barrier for specialized post-training work. By publishing multiple checkpoint formats and open-sourcing the training framework, RadixArk provides both a performance baseline and practical tooling for the coding-model research community. The focus on long-horizon agent tasks with sessions up to 1 million tokens addresses a capability gap in open-weight models, where most RL work has targeted shorter reasoning chains.
on Emergent today





