
DaisyRose-27B: RadixArk's First Model Doubles Terminal-Bench to 12.0%
By Ali Al-Montaser — Senior AI Technical Editor
RadixArk launched DaisyRose-27B, built on Qwen3.8-27B with Apache 2.0 license and 256K context. Trained on 1,200 agentic tasks in 18 days on 144 H200 GPUs, it doubled Terminal-Bench 4.1 from 5.6% to 12.0% and cut training cost by 57%. Available now on Hugging Face with SGLang and DSpark support.
Executive Overview
RadixArk has officially announced the launch of its first model, DaisyRose-27B, built on Qwen3.8-27B under the Apache 2.0 license and available for immediate download on Hugging Face. The model was trained on 1,200 agentic tasks over 18 days on 144 H200 GPUs, achieving a jump in Terminal-Bench 4.1 from 5.6% to 12.0% and reducing training cost by 57%. This release marks a significant milestone in open-source agentic AI, offering developers a free, locally runnable model with a 256K context window and advanced reinforcement learning capabilities.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Completely free — Apache 2.0 open-source license, no licensing fees or subscriptions. Operating costs are limited to your own GPU hosting. |
| 🌐 Platforms & Immediate Availability | Hugging Face (direct download) + SGLang cookbook for running on port 30000 + DSpark speculative decoding support + Miles training framework. |
| ⚡ Performance & Speed Benchmarks | Terminal-Bench 4.1: 5.6% → 12.0% | Terminal-Bench 2.1: 73.0 → 75.5 | DeepSWE v1.1: 42.2 → 46.2 | SWE Bench Pro v2: 91.3 → 94.7 | HLE: 33.9 → 34.3 | IFBench: 79.5 → 82.4 | GPQA Diamond: 89.2 → 90.7 |
| 🛡️ Security & Breach Resistance | Isolated sandbox environments with 99% success rate (exceeded 99.95% since late October), 944,000 sandboxes launched with a maximum of 1,566 concurrent. |
| 🧠 Context Window | 256K tokens (inherited from Qwen3.8-27B) + agentic training sessions up to 1M tokens via harness summarization. |
| 🌍 Arabic & Regional Support | Inherits Qwen3.8-27B's multilingual tokenizer and image understanding. No explicit Arabic evaluation in the official statement. |
Deep-Dive Features & Architecture
DaisyRose-27B is built on the Qwen3.8-27B architecture, fully retaining its tokenizer, chat template, image understanding, and 256K token context window. The model underwent reinforcement learning (RL) using RadixArk's open-source Miles framework in a fully asynchronous setup: 128 H200 GPUs run inference on agentic tasks while 16 GPUs run the trainer, with model updates after each step. Training leveraged 1,200 curated tasks selected in collaboration with Eigent AI from thousands of candidates, each targeting a specific failure mode in the base model and requiring over 100 tool calls and 256K output tokens in sessions lasting up to 1.5 hours.
The training infrastructure achieved an MFU efficiency of 37% compared to 16% for the open-source Megatron-based trainer, resulting in a 57% reduction in training cost for the same task. Training speed reached 16,800 tokens/second across 144 H200 GPUs versus 8,300 for the open-source trainer (2x speed). Training ran for weeks totaling 63,000 GPU hours with consistent speed and efficiency from the first step to the last. Additionally, cloud infrastructure costs for sandboxes were reduced by 80% compared to common cloud providers.
Benchmark & Competitive Performance
DaisyRose-27B outperforms Qwen3.8-27B on all seven reported benchmarks. The largest jump is in Terminal-Bench 4.1, where pass@1 rose from 5.6% to 12.0% (more than double). In SWE Bench Pro v2 it increased from 91.3 to 94.7, in IFBench from 79.5 to 82.4, in GPQA Diamond from 89.2 to 90.7, and in DeepSWE v1.1 from 42.2 to 46.2. Scores are calculated as the highest result between BF16 and FP8 runs, with a difference of less than one point where dual evaluation was performed. On mixture-of-experts models, the trainer achieved even larger gains: 6.2× on Qwen3.6-35B-A3B and over 10× on GLM-5.3-Flash (approximately 300 billion parameters).
Industry Impact & Enterprise Adoption
DaisyRose-27B represents a practical opportunity for developers worldwide, including Arabic-speaking developers, because it is free under the Apache 2.0 license and can be run locally on a single GPU architecture, eliminating monthly subscription costs for startups. The inherited multilingual tokenizer from Qwen3.8-27B provides basic Arabic coverage, but the official statement does not provide specific Arabic token efficiency figures, so internal evaluation is recommended before production adoption. The 256K context window enables processing of long Arabic documents (reports, contracts, research) in a single session, while the extended agentic capabilities (1M training tokens) open use cases in software task automation and analysis. The model's availability on Hugging Face with SGLang and DSpark support lowers the barrier to entry for enterprises looking to deploy agentic AI without vendor lock-in.
Conclusion
RadixArk's DaisyRose-27B is a noteworthy advancement in open-source agentic AI, delivering substantial benchmark improvements and cost efficiencies. With its Apache 2.0 license, 256K context, and immediate availability on Hugging Face, it empowers developers to build and deploy sophisticated AI agents locally. While Arabic-specific performance metrics are not yet published, the model's multilingual foundation and free access make it a compelling option for global and regional developers alike.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
The model is completely free and open-source under the Apache 2.0 license. It can be downloaded directly from Hugging Face with no licensing fees or subscriptions. Operating costs are limited to your own GPU hosting via SGLang.
It is available for immediate download on Hugging Face on the RadixArk page, with an SGLang cookbook guide for server configuration on port 30000. No managed cloud API version has been announced by RadixArk yet.
The model retains the 256K token context window inherited from Qwen3.8-27B, with support for extended agentic training sessions up to 1M tokens via harness summarization.
DaisyRose-27B outperforms Qwen3.8-27B on all reported benchmarks: Terminal-Bench 4.1 from 5.6% to 12.0%, Terminal-Bench 2.1 from 73.0 to 75.5, SWE Bench Pro v2 from 91.3 to 94.7, and GPQA Diamond from 89.2 to 90.7.
The official statement does not mention an explicit Arabic evaluation, but the model inherits the multilingual tokenizer and image understanding of Qwen3.8-27B, providing basic Arabic coverage. Exact Arabic token efficiency figures are not available in the source.

AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
