← Back to BLACKWIRE EMBER BUREAU MODEL EFFICIENCY Diagram of DaisyRose-27B training pipeline showing data curation, asynchronous Miles framework, and sandbox execution.

RadixArk’s training loop: 128 inference GPUs, 16 trainer GPUs, and a high‑throughput sandbox service.

DAISYROSE-27B DOUBLES TERMINAL‑BENCH PASS RATE WITH 1,200 CURATED TASKS

*RadixArk’s DaisyRose-27B pushes post‑training efficiency to new heights. The model lifts Terminal‑Bench 4.1 pass@1 from 5.6% to 12.0% while slashing GPU cost by 57%. The gains stem from a tight data pipeline, an asynchronous training backend, and a home‑grown sandbox service.*

By EMBER Bureau - BLACKWIRE  |  November 19, 2026, 05:00 CET  |  DaisyRose-27B, post-training efficiency, Terminal-Bench, reinforcement learning, sandbox infrastructure

RadixArk has unveiled DaisyRose-27B, its first post‑training model built on Qwen3.8-27B and refined with the Miles reinforcement‑learning framework. In just 18 days on 144 H200 GPUs, the team ran 1,200 curated long‑horizon tasks, some stretching to a million tokens, to reshape the model’s error profile. The result is a double‑digit lift on the notoriously hard Terminal‑Bench 4.1, moving pass@1 from 5.6% to 12.0%, and consistent gains on a suite of reasoning and coding benchmarks. The release includes BF16, FP8 and NVFP4 checkpoints, inviting developers to test and extend the model immediately.

What performance gains does DaisyRose-27B deliver?

DaisyRose-27B, built on Qwen3.8-27B, outperforms its predecessor on every benchmark above a score of 1. The standout is Terminal‑Bench 4.1, where pass@1 jumps from 5.6% to 12.0%. Gains also appear across a range of reasoning, instruction‑following, and coding tests. The model ships in BF16, FP8 and NVFP4 formats, inviting immediate community experimentation. RadixArk attributes the leap to three core ingredients: high‑quality data, an efficient training backend, and robust sandboxes.

How did high‑quality data drive the improvement?

RadixArk partnered with Eigent AI to sift through thousands of candidate tasks. Only about 1,200 made the final set after rigorous filtering. Each task targets a specific mistake made by the original Qwen3.8-27B, and is tuned until its solve rate sits in a learning sweet spot—neither trivial nor impossible. Tasks chain multiple interdependent problems, forcing agents to plan, reason, and execute over up to 1.5 hours, with more than 100 tool calls and over 256 k output tokens per attempt. Every task was reviewed alongside model attempts; any that could be passed without genuine solving were removed, eliminating reward‑hacking avenues.

Our magic lies in the following three key ingredients: high‑quality training data, efficient training backend, and robust sandboxes.

What efficiency breakthroughs cut training cost by more than half?

Training ran on RadixArk’s Miles framework in fully asynchronous mode. While 128 H200 GPUs handled inference on agent tasks, 16 additional GPUs learned from the trajectories. This pipeline achieved 57% lower training cost, with 37% of each GPU’s peak compute devoted to useful model FLOPs (MFU) versus 16% on comparable open‑source trainers. Token throughput hit roughly 16,800 tokens per second across the 144 GPUs, double the 8,300 tokens per second of the open‑source baseline. The run consumed 63,000 GPU‑hours over nine weeks, and speed held steady throughout.

How reliable was the sandbox infrastructure at scale?

Every agent attempt executed inside a fresh sandbox built on the open‑source AgentENV project. Between mid‑September and mid‑November the service launched about 944,000 sandboxes, peaking at 1,566 concurrent instances. Infrastructure grew from a single machine to twelve, with ten upgrades deployed during training. Success rates stayed above 99% for sandbox creation, climbing to over 99.95% after late October. Memory‑elastic allocation kept costs down, delivering roughly an 80% cost advantage over popular cloud sandbox providers.

Frequently asked questions

What is the pass@1 improvement on Terminal‑Bench 4.1?

DaisyRose-27B raises pass@1 from 5.6% to 12.0% on Terminal‑Bench 4.1.

How many GPUs and how much compute time were used to train DaisyRose-27B?

The model was trained on 144 H200 GPUs for a total of 63,000 GPU‑hours over nine weeks.

What cost advantage does the sandbox service provide?

The cloud infrastructure cost for sandboxes was about 80% lower than running the same sandboxes with popular cloud providers.

DaisyRose-27B proves that strategic data curation and purpose‑built infrastructure can deliver outsized performance without a massive compute bill. By publishing checkpoints and the Miles framework, RadixArk opens the door for other teams to replicate the efficiency gains. The model’s rapid ascent on long‑horizon benchmarks signals a shift toward leaner, more capable AI systems—an evolution that could reshape how research labs allocate scarce GPU resources.

Sources: [Hacker News – DaisyRose-27B blog post on radixark.com]