RadixArk has released DaisyRose-27B, its first open-weight model: an agentic coding model built from Qwen3.8-27B and trained exclusively with reinforcement learning, with no supervised fine-tuning or distillation. The weights are available under Apache 2.0 and can directly replace the base model.
The most striking result concerns Terminal Bench 4.1: pass@1 rises from 5.6% to 12.0%, more than doubling. Across the seven benchmarks the company reports, DaisyRose beats the base model on all of them, with smaller gains on easier coding tests: +4.0 points on DeepSWE v1.1, +3.4 on SWE Bench Pro v2 and +2.5 on Terminal Bench 2.1. Knowledge gains are modest: +0.4 on text-only HLE and +1.5 on GPQA Diamond. All scores are self-reported by RadixArk and have not yet been checked by independent evaluations.

Image: RadixArk — https://huggingface.co/RadixArk/DaisyRose-27B
How it was trained
RadixArk built about 1,200 tasks, selected from thousands of candidates, each designed to target a specific failure of the base model. Each sandbox attempt can run up to an hour and a half, with more than 100 tool calls and over 256,000 output tokens; agent sessions reach up to one million tokens.
Training ran for 18 days on 144 H200 GPUs in fully asynchronous mode: 128 handled agent inference and 16 handled training, for a total of 63,000 GPU-hours. The company says its internal backend reaches 37% compute efficiency, versus 16% for an open-source Megatron-based trainer, and cuts GPU time by 57%. The framework is Miles, RadixArk's reinforcement learning library, published on GitHub.
The sandboxes were created about 944,000 times between mid-September and mid-November, with a peak of 1,566 in parallel and a 99% success rate. RadixArk puts the cost at 80% below popular cloud sandbox providers.
Size and use
The model has 27.8 billion dense parameters, a 262,144-token context window, and accepts text and images as input while producing text. It runs on one GPU: about 52 GB in BF16, 29 GB in FP8 and 23 GB in NVFP4, the latter only on Blackwell GPUs. GGUF files are also available. Weights are published on Hugging Face and ModelScope.

Image: miles.radixark.com — https://miles.radixark.com
The gaps with larger models remain significant. MiMo V2.6 Flash, with 309 billion total parameters, scores 16.1 on Terminal Bench 4.1, above DaisyRose. Inkling 2 Small, with 594 billion total, leads on DeepSWE and HLE. In RadixArk's comparison of nine models, DaisyRose is first only on IFBench, with 82.4 points.
What remains to be verified
The most interesting point for teams training specialized agents is cost: if the efficiency figures hold outside the company's tests, smaller teams could afford post-training of this kind. RadixArk now offers this approach as a service. For now, however, the numbers come from the company that built the model, which has asked users to report where it falls short.

No comments yet. Be the first!