deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

RadixArk's DaisyRose-27B more than doubles Terminal-Bench score with 1,200-task RL run

RadixArk says reinforcement learning on just 1,200 curated long-horizon tasks lifted DaisyRose-27B from 5.6% to 12.0% on Terminal-Bench 4.1, in an 18-day run on 144 H200 GPUs.

RadixArk's DaisyRose-27B more than doubles Terminal-Bench score with 1,200-task RL run

RadixArk has released DaisyRose-27B, its first model, and claims that a deliberately small reinforcement learning run more than doubled the base model's score on a difficult agentic benchmark. According to the company's announcement, training Qwen3.8-27B on just 1,200 curated long-horizon tasks raised its Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. The run took 18 days on 144 H200 GPUs, totalling roughly 63,000 GPU-hours, and RadixArk is publishing BF16, FP8 and NVFP4 checkpoints on Hugging Face.

Fewer tasks, chosen against the model's failures

RadixArk built the training set with Eigent AI, starting from a large pool of tasks spanning machine learning, software engineering and science. Thousands of candidates were generated; about 1,200 made the final cut, selected against three criteria.

First, every task targets a mistake the original model actually makes. The team ran Qwen3.8-27B on each candidate and redesigned tasks to surface the specific errors they wanted trained away, then tuned difficulty so each task's solve rate landed in a band that was neither trivial nor hopeless. Second, tasks are long jobs that chain several interdependent problems, forcing multi-stage reasoning, planning and execution; a single attempt can run up to 1.5 hours in a sandbox with more than 100 tool calls and more than 256k output tokens. Third, the team manually reviewed tasks alongside model attempts and removed or fixed anything that could be passed without genuinely solving it, an explicit defence against reward hacking.

Asynchronous training infrastructure

Training used Miles, RadixArk's open-source reinforcement learning framework for post-training, in fully asynchronous mode: the agent keeps working while the trainer learns, and neither waits for the other. During the release run, 128 H200 GPUs handled agent inference while 16 ran the trainer, with updated weights pushed to the inference fleet after every step. The internal backend is built for agent workloads, learning from a handful of very long sessions per step, some reaching 1M tokens, folded into multiple trajectories through harness summarization.

The efficiency claims are the most striking part. RadixArk reports 37% model FLOPs utilization on its backend versus 16% for the open-source trainer, which is built on Megatron, on comparable runs, implying the same training needs 57% less GPU time. End to end, including agent time, the run processed about 16,800 tokens per second across the 144 GPUs versus roughly 8,300 for the open-source trainer. The company says gains are larger on mixture-of-experts models: 6.2x on Qwen3.6-35B-A3B and more than 10x on the roughly 300B-parameter GLM-5.3-Flash.

Sandboxes at scale

Every training attempt executes in a fresh sandbox environment where the agent runs commands and edits files before being graded. RadixArk ran these on its own service, built on the open-source AgentENV project, scaling from one machine to 12. Between mid-September and mid-November the service ran about 944,000 sandboxes for this project, peaking at 1,566 running concurrently. Ninety-nine percent of sandbox creations succeeded overall, a figure above 99.95% since late October, and ten upgrades were shipped, nearly all while training was live. RadixArk estimates its sandbox cloud costs came in about 80% below popular commercial sandbox providers, partly by letting each sandbox's memory grow and shrink with actual usage.

Evaluation caveats

RadixArk says DaisyRose-27B also beats the original Qwen3.8-27B on every other benchmark it tested, covering reasoning, instruction following and coding. But all numbers are self-reported, and the methodology matters: Terminal-Bench 4.1 results are averages over 8 runs per task using RadixArk's internal mini-swe-agent harness with summarization enabled, so scores from other harnesses may not be directly comparable. No independent evaluation accompanies the release.

Why it matters

This is a notable data point for the argument that RL post-training rewards task quality over quantity. A set of 1,200 hand-curated tasks, an 18-day run and a 144-GPU cluster are modest by frontier standards, yet RadixArk claims a 2.1x improvement on a benchmark designed to be brutally hard for agents. The absolute score, 12%, remains low, but the claimed recipe, targeting tasks at known model failures, keeping the trainer and agent asynchronous, and running cheap dedicated sandboxes, offers a template other teams could copy. If the efficiency and cost figures hold up under outside scrutiny, the release lowers the barrier for organizations doing specialized post-training on open models, and the published checkpoints plus the open-source Miles framework make the approach reproducible in principle.

  • #reinforcement-learning
  • #open-source
  • #language-models
  • #benchmarks
  • #agents

Related posts