LLMs

RadixArk releases DaisyRose-27B, a 27-billion-parameter model optimized for long-horizon reasoning tasks

DaisyRose-27B doubles Terminal-Bench 4.1 performance with post-training on 1,200 curated tasks, trained in 18 days on 144 H200 GPUs.

Last verified:

DaisyRose-27B: Scaling Post-Training with Task Curation

According to Hugging Face Blog, RadixArk released DaisyRose-27B, a 27-billion-parameter model built by applying reinforcement learning to Qwen 3.8-27B. The standout feature is not model scale but training efficiency: the team curated only 1,200 tasks—sourced in partnership with Eigent AI—and completed the full post-training cycle in 18 days across 144 H200 GPUs. The model maintains the original Qwen architecture, tokenizer, 256K-token context window, and multimodal image understanding, but gains significant reasoning and coding capabilities through targeted fine-tuning on long-horizon agent tasks with sessions spanning up to 1M tokens.

Benchmark Gains Across Reasoning and Coding

The largest performance jump appears on Terminal-Bench 4.1, where pass@1 more than doubles from 5.6% to 12.0%. Beyond that headline figure, Hugging Face Blog reports consistent improvements across multiple evaluation suites: DeepSWE v1.1 rises from 42.2 to 46.2 percentage points, SWE-Bench Pro v2 from 91.3 to 94.7, IFBench from 79.5 to 82.4, and GPQA Diamond from 89.2 to 90.7. Terminal-Bench 2.1 shows a more modest 2.5-point lift (73.0 to 75.5), and HLE text-only gains only 0.4 points, suggesting diminishing returns on tasks where the base model already performs well. BF16 and FP8 checkpoint runs differ by less than 1 percentage point on evaluated benchmarks, indicating stable performance across precision formats.

Deployment and Open-Weights Release

The model is available under Apache 2.0 licensing in three formats: BF16, FP8, and NVFP4. Hugging Face Blog recommends serving DaisyRose-27B with SGLang and DSpark speculative decoding to maximize inference throughput on GPU infrastructure. The model defaults to chain-of-thought reasoning, returning thinking steps in a separate reasoning_content field—useful for auditing agentic workflows but requiring application-level handling of the dual-output structure. Deployment guidance and launch configurations for different GPU SKUs are provided in the SGLang cookbook.

Task Quality as a Bottleneck

The training methodology highlights an industry shift: post-training efficiency now trades on data quality, not volume. Eigent AI and RadixArk built task candidates using three selection criteria—each targeting observed failure modes of the original model, targeting diversity across problem domains, and meeting a minimum difficulty threshold. This curation process distilled thousands of candidates down to 1,200 final tasks. The compressed training schedule (18 days vs. typical weeks-to-months post-training) suggests that adversarial alignment against known weaknesses may scale more efficiently than broad instruction-following data.

Why This Matters

Teams evaluating open-weights models for agentic reasoning pipelines now have a concrete data point on post-training ROI: substantial reasoning gains (2.14× on Terminal-Bench 4.1) are achievable without full-scale retraining, using modest compute (144 H200s for 18 days) and curated task sets rather than internet-scale datasets. For organizations building software engineering agents or long-horizon planning systems, DaisyRose-27B offers a licensed alternative to closed-API reasoning models, with inference costs anchored to H100/H200 hardware rather than vendor pricing. The Apache 2.0 release also de-risks vendor lock-in for proprietary deployments. However, the Terminal-Bench 4.1 scores (12.0%) remain below commercial reasoning models—the meaningful test will be whether independent reproduction confirms these gains and whether downstream agent success rates (task completion on unseen domains) validate the benchmark improvements.

Frequently Asked Questions

What is DaisyRose-27B and how does it differ from Qwen 3.8-27B?

DaisyRose-27B is a post-trained version of Qwen 3.8-27B fine-tuned using reinforcement learning on long-horizon agent tasks. It retains the base model's 256K-token context, tokenizer, and architecture, but achieves higher performance on reasoning and coding benchmarks through specialized task-based training.

How efficient is the training process?

According to Hugging Face Blog, DaisyRose-27B was trained on just 1,200 curated tasks in 18 days using 144 H200 GPUs, demonstrating a compact post-training regime optimized for data quality over data scale.

What benchmarks show the largest improvements?

Terminal-Bench 4.1 shows the largest gain, with pass@1 rising from 5.6% to 12.0%. Gains across other benchmarks range from 0.4 to 4.5 percentage points.

How can I run DaisyRose-27B?

The model is available under Apache 2.0 in BF16, FP8, and NVFP4 formats. Hugging Face recommends serving it with SGLang and DSpark speculative decoding, with example configurations provided in the SGLang cookbook.

#open-weights #post-training #reasoning #agentic-ai #efficiency