RadixArk’s first model doubles Qwen3.8-27B on Terminal Bench 4.1
RadixArk's DaisyRose-27B, an Apache-2.0 reinforcement-learning derivative of Qwen3.8-27B, reports 12.0 on Terminal Bench 4.1 against 5.6 for its base model.

RadixArk has published DaisyRose-27B on Hugging Face, its first model. It is a reinforcement-learning derivative of Alibaba’s Qwen3.8-27B, and its card reports a Terminal Bench 4.1 score of 12.0, up from 5.6 for the base model.
A 27B model trained on agent failures
The model card describes DaisyRose-27B as a multimodal reasoning model: it takes text and images and outputs text. It was trained with reinforcement learning only, according to the card, “without any supervised fine-tuning or distillation.” The card lists 27.8B parameters in total, including a 0.46B vision encoder and a 0.42B multi-token-prediction head. Both of those parts are unchanged from Qwen3.8-27B. The context window is 262,144 tokens.

The training data is about 1,200 long-horizon agent tasks built with Eigent AI. RadixArk says each task targets a mistake the base model makes. An attempt can run up to 1.5 hours in a sandbox, with more than 100 tool calls and more than 256K output tokens. The run used 128 H200 GPUs to serve the agent and 16 more to train, for 18 days on 144 H200s, roughly 63,000 GPU-hours. RadixArk used its own open-source Miles framework for the asynchronous training.
The weights are released under Apache-2.0 in BF16, FP8 and NVFP4. The card gives file sizes of about 52 GB, 29 GB and 23 GB respectively, and says NVFP4 needs a Blackwell GPU. Separate repositories hold the FP8, NVFP4 and GGUF versions, plus a draft model for speculative decoding. The card recommends serving with SGLang and says the model runs on a single GPU, from an H200 down to an RTX 5090 or a DGX Spark. The Hub page also shows vLLM and Transformers snippets, which are generic Hub templates.
Big relative gains, but not the top of its own table
The card’s headline claim is the Terminal Bench 4.1 jump from 5.6 to 12.0. Other rows move less. Against Qwen3.8-27B, the card reports 75.5 versus 73.0 on Terminal Bench 2.1, 46.2 versus 42.2 on DeepSWE v1.1, 94.7 versus 91.3 on SWE Bench Pro v2, 34.3 versus 33.9 on HLE (text only), 82.4 versus 79.5 on IFBench and 90.7 versus 89.2 on GPQA Diamond.
Terminal Bench 4.1: 12.0 for DaisyRose-27B, 5.6 for Qwen3.8-27B
The same table shows DaisyRose is not the strongest model in it. On Terminal Bench 4.1, the card lists MiMo V2.6 Flash at 16.1, GPT-6 Luna at 15.2 and Inkling 2 Small at 13.0, all above 12.0. DaisyRose leads only on IFBench. Inkling 2 Small is also ahead on DeepSWE, SWE Bench Pro v2, HLE and GPQA Diamond. The size gap matters here: the card gives Inkling 2 Small 594B total parameters and MiMo V2.6 Flash 309B, against 27B for DaisyRose.
How the scores compare with other Qwen3.8-27B work
Numbers in context: Terminal Bench scores depend heavily on the harness and benchmark version, so the card’s figures cannot be set directly against other reports. RadixArk used an internal mini-swe-agent harness for 4.1 and Terminus 2 for 2.1, and it reports the higher of its BF16 and FP8 runs, which it says differ by under one point. Competitor numbers come from their developers or third-party leaderboards, and the card says several rows mix harnesses and, for some models, an earlier benchmark version.
Earlier work on the same base model shows a similar pattern. Our report on harness adapters for Qwen3.8-27B covered a small LoRA adapter that lifted the model from 39% to 55% on Terminal-Bench. That is a different route (an adapter, not a full reinforcement-learning run) and a different measurement, so the two results are not comparable. Still, both suggest Qwen3.8-27B has room to improve on agentic terminal tasks. Our weekly roundup of new models lists Qwen4 Max among this week’s releases, so the 27B generation is already being followed by a larger successor from Alibaba.
What the card leaves unverified
Every DaisyRose number comes from RadixArk’s own evaluation setup, and nothing on the card is independently reproduced. The card gives no reward design, learning-rate details or sample of the 1,200 tasks, so the training recipe cannot be reproduced from it. The comparison set also includes models with very different sizes and harnesses, and the card’s Hub page lists a separate blog post by RadixArk that this report did not draw on.
For developers, the practical point is availability: the weights are Apache-2.0 and run on a single GPU, with quantized versions for consumer hardware. Whether the Terminal Bench 4.1 gain carries over to real coding agents is a question third-party testing will have to answer.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



