model release

RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%

TL;DR

RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.

3 min read
0

RadixArk released DaisyRose-27B on November 18, 2026, a 27B-parameter model that raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% over its base, Qwen3.8-27B. It is RadixArk's first model. Checkpoints are available under Apache 2.0 in BF16, FP8 and NVFP4.

The model keeps Qwen3.8-27B's architecture, tokenizer, chat template, image understanding and 256K-token context window. It thinks before answering by default, returning reasoning in reasoning_content and the final answer in content. API pricing is not yet disclosed.

Benchmarks (company-reported)

RadixArk reports the following scores, using the higher of the BF16 and FP8 runs for each benchmark. According to RadixArk, the two differ by less than one point wherever both were evaluated.

Benchmark Qwen3.8-27B DaisyRose-27B
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 (text-only) 33.9 34.3
IFBench 79.5 82.4
GPQA Diamond 89.2 90.7

The largest gain is on Terminal Bench 4.1. The model card includes comparison sources and evaluation settings.

Training setup

RadixArk trained the model with Miles, its open-source reinforcement learning framework, in fully asynchronous mode. It used 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs: 128 ran agent inference and 16 ran the trainer. Sessions ran up to 1M tokens, spanning multiple trajectories through harness summarization. The release run totaled 63,000 GPU-hours.

The tasks were built with Eigent AI, starting from its pool of tasks in machine learning, software engineering and science. Thousands of candidates were built, and about 1,200 made the final set. RadixArk says each task targets a mistake Qwen3.8-27B makes, requires several interdependent steps, and was reviewed to ensure it rewards actually solving the problem. Each attempt gets up to 1.5 hours in a sandbox, with more than 100 tool calls and more than 256K output tokens.

RadixArk makes these efficiency claims about its internal training backend versus the open-source Megatron-based trainer:

  • MFU: 37% on this release, versus 16% for the open-source trainer on comparable runs. RadixArk says this means 57% less GPU time for the same training.
  • Throughput: about 16,800 tokens per second across 144 GPUs, versus about 8,300 for the open-source trainer.
  • MoE models: 6.2x on Qwen3.6-35B-A3B and more than 10x on the roughly 300B-parameter GLM-5.3-Flash.

Sandbox infrastructure

Sandboxes run on a service built on the open-source AgentENV project. Between mid-September and mid-November it ran about 944,000 sandboxes for this project, with up to 1,566 running concurrently. RadixArk says 99% of creations succeeded across nine weeks, and the rate has exceeded 99.95% since late October. It also claims cloud costs about 80% lower than popular cloud sandbox providers, partly by letting each sandbox's memory allocation scale with actual usage.

Running it

RadixArk recommends serving with SGLang and DSpark speculative decoding. The SGLang cookbook has launch configurations by GPU and checkpoint format. The model card's quickstart uses temperature 1.0, top_p 0.95, top_k 20 and min_p 0.0 through an OpenAI-compatible endpoint.

What this means

DaisyRose-27B is an Apache 2.0 agentic-coding variant of an existing open model, with the same context window and tokenizer as its base, so it fits existing Qwen3.8-27B deployments. The reported gains are largest on terminal-based agent work, the target of the training tasks. They are smaller on HLE (text-only), where the score moves from 33.9 to 34.3. All figures are reported by RadixArk, and it invites users to share results, including failures, in the model's Community tab. Separately, the release doubles as a showcase for Miles and RadixArk's training backend, whose efficiency and sandbox figures are the company's own measurements.

Related Articles

model release

Reson8 releases 500M-parameter Rumble-2-Flash speech model as free Apache 2.0 open weights

Amsterdam-based Reson8 (Resonate Labs B.V.) has released Rumble-2-Flash, a 500 million-parameter transcription model covering more than 40 languages, as a free Apache 2.0 open-weights download. The company claims a 7.09% average word error rate across seven languages on Hugging Face's Open ASR Leaderboard multilingual tests.

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use

Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

Comments

Loading...