Skip to content
important New models

RadixArk releases open-weight model DaisyRose-27B for agentic tasks

clearly official source

RadixArk has released the DaisyRose-27B model under the Apache 2.0 license, a fine-tuned version of the Qwen3.8-27B model focused on long agentic tasks. On the Terminal Bench 4.1 benchmark it improved pass@1 from 5.6% to 12.0%, and other coding and instruction-following tests also showed improvement.

RadixArk released the DaisyRose-27B model, its first model, built on the base of the Qwen3.8-27B model and fine-tuned using reinforcement learning with the Miles framework. The model focuses on long agentic tasks with sequences of up to 1 million tokens. Training was conducted on 1 200 carefully selected tasks over 18 days on 144 H200 GPUs. The tasks were created in collaboration with Eigent AI, with roughly 1 200 tasks selected from thousands of candidates for the final training set, targeting specific errors of the original Qwen3.8-27B model and requiring multiple sequential steps.

On the Terminal Bench 4.1 benchmark, pass@1 success improved from 5.6% to 12.0%. Other tests also showed improvement: Terminal Bench 2.1 from 73.0% to 75.5%, DeepSWE v1.1 from 42.2% to 46.2%, SWE Bench Pro v2 from 91.3% to 94.7%, HLE (text-only) from 33.9% to 34.3%, IFBench from 79.5% to 82.4%, and GPQA Diamond from 89.2% to 90.7%. The model retains the architecture, tokenizer, chat template, and 256K-token context window of the Qwen3.8-27B model. The weights are available under the Apache 2.0 license.

According to RadixArk, a custom training backend achieved 37% GPU compute utilization (MFU) compared to 16% for an open trainer built on Megatron, which according to the company means 57% lower GPU time consumption for the same training and twice the speed in terms of tokens per second. In total, training consumed 63 000 GPU-hours on 144 H200 GPUs over several weeks. For the sandboxes in which the agent operates during training, the company used the open-source project AgentENV; between mid-September and mid-November, approximately 944 000 sandboxes were run in it with a creation success rate of over 99%, and according to the company, with cloud infrastructure costs roughly 80% lower than typical commercial cloud sandbox providers.

What changed

Why it matters

This is a freely available model under the Apache 2.0 license, built on the Qwen3.8-27B model, which can be run locally or on your own infrastructure and deployed for agentic and coding tasks without licensing fees. Measurable improvement on benchmarks such as Terminal Bench and SWE Bench Pro shows a concrete shift in the ability to handle long, sequential tool-use tasks, which is relevant for teams building their own coding agents or testing alternatives to closed-source models.

Release card

DaisyRose-27B

RadixArk

open weights
Context
256K tokens
Inputs
text and images as input, text as output
Licence
Apache 2.0
Availability
The weights are available for download on Hugging Face; deployment is recommended using the SGLang tool with DSpark speculative decoding.
Documented measurements
  • Terminal Bench 4.1 pass@1 12,0 % (oproti 5,6 % u modelu Qwen3.8-27B) The model improves the success rate of completing terminal agentic tasks compared to the base model.
  • Terminal Bench 2.1 75,5 % (oproti 73,0 % u modelu Qwen3.8-27B) Shows improvement on an older version of the terminal agentic benchmark.
  • DeepSWE v1.1 46,2 % (oproti 42,2 % u modelu Qwen3.8-27B) Shows improvement in software engineering capabilities.
  • SWE Bench Pro v2 94,7 % (oproti 91,3 % u modelu Qwen3.8-27B) Shows improvement in solving software tasks on this benchmark.
  • HLE (text-only) 34,3 % (oproti 33,9 % u modelu Qwen3.8-27B) Shows a small improvement on a test of difficult knowledge questions.
According to the sources, it is suitable for
  • Long agentic tasks with sequences of up to 1 million tokens
  • Terminal and coding agentic tasks requiring multiple sequential steps
  • Tasks requiring precise instruction-following

According to the source, the DaisyRose-27B model achieves a higher score than the base model Qwen3.8-27B on all the listed benchmarks.

The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →

Two audiences, two different impacts

What this means

01

For individuals

Developers and people experimenting with AI agents for programming are getting a new open model available for download and local deployment, with an architecture and 256K-token context window identical to the Qwen3.8-27B model.

What to do Try the DaisyRose-27B model locally (e.g. via SGLang) for agentic and coding tasks.
More practical updates →
02

For a business

Companies developing or deploying agentic AI tools for coding and task completion are getting another open alternative under the Apache 2.0 license, which can be deployed without the licensing fees of closed-source providers.

Development
What to decide Consider trying the DaisyRose-27B model as an open-weight alternative for internal agentic/coding tools, thanks to its Apache 2.0 license.
More business impacts →
DaisyRose-27B Eigent AI Miles Qwen3.8-27B RadixArk SGLang

Check the original

Event sources

clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
Hugging Face Blog primary source · first detected DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency