RadixArk Launches DaisyRose-27B: First Open Model Doubles Terminal-Bench Score in 18 Days
RadixArk releases DaisyRose-27B, built on Qwen3.8-27B with 1,200 curated RL tasks on 144 H200 GPUs, raising Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.

By Clara Tan, Artificial Intelligence
, updated 4 min read
RadixArk ships its first model
RadixArk, the infrastructure startup founded by core maintainers of the SGLang project, has released DaisyRose-27B, its first open-weight model. Announced on 18 November 2026, the model is built on Qwen3.8-27B and trained with reinforcement learning using Miles, RadixArk's open-source RL framework.
Training details
Base model: Qwen3.8-27B
Framework: Miles (open-source RL framework combining fast multi-turn SGLang rollouts, fully asynchronous RL, and Token In Token Out)
Compute: 144 H200 GPUs for 18 days
Training data: 1,200 curated long-horizon agent tasks across machine learning, software engineering, science and other domains
Session length: Up to 1 million tokens per agent session
Checkpoints: BF16, FP8 and NVFP4 formats released on Hugging Face
Benchmark results
Benchmark | Qwen3.8-27B | DaisyRose-27B | Gain |
|---|---|---|---|
Terminal-Bench 4.1 pass@1 | 5.6% | 12.0% | +114% |
IFBench (instruction following) | , | Higher | , |
GPQA Diamond (reasoning) | , | Higher | , |
Other coding/reasoning benchmarks | , | Higher across the board | , |
The Terminal-Bench improvement is the standout result: a more than doubling of pass@1 on a notoriously difficult long-horizon agent benchmark, achieved with a relatively small curated dataset and modest compute budget.
Why it matters
RadixArk's approach demonstrates that carefully designed tasks and efficient infrastructure can yield significant post-training gains without massive compute. The company's thesis, that open infrastructure (SGLang for inference, Miles for RL training) should be a shared foundation, is now validated by a shipping model.
The model is available today on Hugging Face:
Open infrastructure thesis
RadixArk was founded by Ying Sheng and Banghua Zhu, core maintainers of the SGLang project. The company raised a $100M seed round at a $400M post-money valuation in May 2026 to build open infrastructure for frontier AI. Their platform combines SGLang for high-throughput inference and Miles for RL training, offered as managed services.
The DaisyRose-27B release proves the loop: open inference engine (SGLang) enables fast rollouts for RL training (Miles), which produces better models that can be served on SGLang. All three components, SGLang, Miles, and now DaisyRose, are open source under Apache 2.0.
Try it
The model runs on standard inference stacks. On Hugging Face you can test it in the browser, or deploy via SGLang, vLLM or any compatible server. RadixArk's managed platform offers production inference and RL training as a service for teams who don't want to run their own infrastructure.
Sources
- 1RadixArk blog announcementradixark.com
- 2Hugging Face blog posthuggingface.co
- 3Hugging Face model page (BF16)huggingface.co
- 4RadixArk X announcementx.com