A doubling, read in absolute terms

The release headline writes itself: reinforcement learning doubled the base model's score. The arithmetic underneath it is narrower. Qwen3.8-27B scored 5.6% pass@1 on Terminal-Bench 4.1. DaisyRose-27B scores 12.0%, a gain of 6.4 points. Both numbers come from RadixArk's own model card, and the second is a genuine improvement on a hard benchmark. It also corresponds to an approximately 88% failure rate across the evaluated attempts.

Context from the same table cuts the claim down further. DaisyRose does not lead Terminal-Bench 4.1 even among the models RadixArk chose to compare it with: MiMo V2.6 Flash is listed at 16.1%, GPT-6 Luna at 15.2% and Inkling 2 Small at 13.0%. Those three figures come from the official leaderboard, measured under their own conditions, while DaisyRose was measured on RadixArk's internal harness, so the table lists them higher without establishing a controlled head-to-head. What the card does establish is narrower: targeted RL moved the base model's reported score from 5.6% to 12.0%, using about 63,000 GPU-hours of compute, without changing the model's architecture at all.

What the training actually bought

The method is the interesting part. There was no supervised fine-tuning and no distillation. RadixArk and Eigent AI built about 1,200 tasks, each aimed at a mistake the base model makes, each a long job of several interdependent problems, and each allowing up to 1.5 hours in a sandbox, more than 100 tool calls and more than 256K output tokens per attempt. RadixArk says it reviewed every task against the model's attempts and fixed or removed any that could be passed without solving the intended problem. Training ran asynchronously for 18 days on 144 H200 GPUs, roughly 63,000 GPU-hours, with 128 GPUs generating agent trajectories while 16 trained on them.

The gains outside the headline benchmark are smaller in the same table. Terminal-Bench 2.1 rose 2.5 points, DeepSWE v1.1 rose 4 points, SWE-Bench Pro v2 rose 3.4 points, IFBench rose 2.9 points, and GPQA Diamond and HLE moved by 1.5 points and 0.4 points. Anyone reading the 12.0% as evidence of a generally smarter 27B model is reading past the rest of RadixArk's own table.

How much weight to put on the numbers

These are the developer's measurements, taken on the developer's harness. Terminal-Bench 4.1 was run on RadixArk's internal mini-swe-agent harness with summarization enabled, averaged over 8 repetitions, and the published DaisyRose figures are the higher of its BF16 and FP8 runs. The card is transparent about all of that, which is more than many releases offer, and it also means an independent run under a different harness should be expected to land somewhere else. The comparison rows for rival models are drawn from leaderboards and other model cards, not from a single controlled run.

The practical offer is clearer than the benchmark. The weights are downloadable under Apache 2.0, the tokenizer and chat template are unchanged, and RadixArk describes the checkpoint as a drop-in replacement for Qwen3.8-27B in an existing stack. The NVFP4 weights take about 23 GB, which is a weight footprint, not a serving-memory figure at a given context length, and that format requires a Blackwell GPU; RadixArk says the model runs on a single GPU, including an RTX 5090 and DGX Spark. For a team that wants a terminal-heavy agent it can run itself, that is testable this week on its own tasks, which is the only evaluation that will settle whether the doubling matters.

Sources

  1. [1] RadixArk, DaisyRose-27B model card on Hugging Face — read first-hand: the RL-only method, the evaluation table and its footnotes, the training data and infrastructure sections, the weights and license.Read source