RadixArk ships DaisyRose-27B, an RL-only 27B agentic coding model
Original title: RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only
Apache 2.0 weights were post-trained from Qwen3.8-27B; RadixArk says Terminal Bench 4.1 pass@1 rose from 5.6 to 12.0.
RadixArk released DaisyRose-27B this week as its first open-weight model, a dense 27.8B thinking model with a 262,144-token context that runs on one GPU. The company says it post-trained Qwen3.8-27B with reinforcement learning only, using about 1,200 tasks, 144 H200 GPUs and 18 days, with no supervised fine-tuning or distillation. Self-reported scores beat the base on seven benchmarks, while harder agentic tests still trail larger mixture-of-experts models.
Unconfirmed: Benchmark scores and training-efficiency figures are self-reported by RadixArk and not independently verified.
- BF16 needs about 52 GB, FP8 about 29 GB, and NVFP4 about 23 GB on Blackwell.
- IFBench at 82.4 led RadixArk’s nine-model table; HLE text-only was 34.3.
- Training used 63,000 GPU-hours and about 944,000 sandboxes, the company says.
- Weights are on Hugging Face in BF16, FP8, NVFP4 and GGUF.
- Vendors
- RadixArk
- Organizations
- Eigent AI
- AI models
- DaisyRose-27BQwen3.8-27B
From the original article
8 min read at marktechpost.com
RadixArk released DaisyRose-27B, an Apache 2. 0 agentic coding model post-trained from Qwen3. 8-27B with reinforcement learning only.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.