Back home

DaisyRose-27B

x+web 0.87
Visit source

Intro

DaisyRose-27B is RadixArk’s first model, released on 18 November 2026: a reinforcement-learning-only post-train of the existing Qwen3.8-27B, with no supervised fine-tuning or distillation. The official post DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency (dated November 18, 2026) says: “Today we are releasing DaisyRose-27B, RadixArk’s first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles.” Weights are on RadixArk/DaisyRose-27B. Hugging Face records the repo as created at 2026-11-18T05:31:35Z. The license is Apache-2.0, with BF16, FP8, and NVFP4 checkpoints.

In this scan slot, @mutianba pointed to the blog at 2026-11-18T21:03:26Z. The repo and blog belong to the official 18 November release, which is earlier than this 08:32 window; earlier slots did not file the name, so this slot backfills the official introduction. The open-source training framework Miles is an existing RadixArk project (radixark/miles was created on 2025-10-09) and is not filed here as a newly emerged standalone product name.

This entry records the model name DaisyRose-27B. Do not treat it as Qwen3.8-27B itself, or as a new version number of the Miles framework.

Advertisement

Explain

The blog says the training tasks were built with Eigent AI, with about 1,200 selected from candidates in machine learning, software engineering, science, and other fields. Selection criteria included targeting mistakes the base model makes and tuning solve rates so tasks were neither too easy nor too hard; chaining interdependent problems into long jobs, with up to 1.5 hours in a sandbox, more than 100 tool calls, and more than 256K output tokens per attempt; and reviewing tasks so they could not be passed by reward hacking. Training ran in Miles’s fully asynchronous mode: 128 H200s ran agent inference and 16 ran the trainer. The release run is described as training for weeks on 144 H200s, 63,000 GPU-hours in total. The post also gives efficiency figures versus an open-source trainer, such as 37% model FLOPs utilization on this run versus 16% for the comparison trainer, and about 2× end-to-end throughput. Those are company claims. The sandbox service is built on open-source AgentENV; the authors say it ran about 944,000 sandboxes over nine weeks, with up to 1,566 running at once.

On results, the largest stated gain is Terminal-Bench 4.1 pass@1, from 5.6% for the base model to 12.0% (averaged over 8 runs per task, using RadixArk’s internal mini-swe-agent harness with summarization on). The model card adds that DaisyRose-27B keeps the base model’s architecture, tokenizer, chat template, image understanding, and 262,144-token context, and is intended as a drop-in replacement for Qwen3.8-27B. It has 27.8B parameters in total, including a 0.46B vision encoder and a 0.42B multi-token-prediction head, both left unchanged from the base model. SGLang documents a deployment recipe. When citing benchmark numbers, attribute them to the 18 November 2026 official blog or model card, and note the evaluation setup and the comparison to the base model.

Usage

  1. Read the official blog at https://www.radixark.com/blog/daisyrose-27b and the Hugging Face model card at https://huggingface.co/RadixArk/DaisyRose-27B . SGLang’s deploy note is at https://docs.sglang.io/cookbook/autoregressive/RadixArk/DaisyRose-27B . The Miles training framework lives at https://github.com/radixark/miles and is not the new name in this entry;
  2. Download the BF16, FP8, or NVFP4 weights as described on the model card, and try them with the recommended inference stack on suitable hardware. Hardware and memory requirements follow the live model card;
  3. When citing the Terminal-Bench 4.1 move from 5.6% to 12.0%, the 1,200 tasks, 144 H200s, or 63,000 GPU-hours, attribute them to the 18 November 2026 official blog or model card and re-check the evaluation setup;
  4. Do not treat DaisyRose-27B as Qwen3.8-27B, and do not treat Miles as a newly emerged standalone product from this slot. This page only restates the official blog and model card, and notes that this is a backfill of the 18 November release.