RadixArk Releases DaisyRose-27B: Pure Reinforcement Learning Doubles Terminal-Bench 4.1 Score on Qwen3.8-27B
RadixArk, the AI infrastructure company behind the open-source SGLang serving engine and Miles reinforcement-learning stack, has released its first open-weight model: **DaisyRose-27B**. Built directly on top of Alibaba's [Qwen3.8-27B](https://www.mushimu.com/post/8976) and published under a permissive Apache 2.0 license in BF16, FP8, and NVFP4 checkpoints, the 27-billion-parameter multimodal reasoning model is an explicit bet that pure reinforcement learning—without a single step of supervised fine-tuning or distillation—can extract frontier-class agentic coding performance from a dense model small enough to run on a single GPU. The release arrives at an inflection point for local and mid-sized open-weight builders. While Alibaba recently introduced its cloud-only [Qwen4-Max](https://www.mushimu.com/post/9581) without open weights for a Qwen 4 27B tier yet, `Qwen3.8-27B` remains the workhorse dense foundation for single-node deployments. Rather than altering the base model's tokenizer, chat template, image-understanding pipeline, or 256K-token context window, RadixArk kept DaisyRose-27B as a strict drop-in replacement—concentrating its entire compute budget on post-training efficiency and long-horizon agentic execution. ### Why Pure RL—and Why Only 1,200 Tasks? Conventional post-training pipelines lean heavily on supervised fine-tuning (SFT) over synthetic trajectories or distillation from a larger proprietary teacher before applying light RL. RadixArk bypassed both stages. Working alongside Eigent AI, the team started with thousands of candidate tasks across software engineering, machine learning, and scientific problem-solving, then winnowed them down to roughly **1,200 curated environments**. Every surviving task had to clear three filters. First, it was redesigned specifically around failures observed when running `Qwen3.8-27B` directly, calibrated so the base model's pass rate landed inside an informative gradient window rather than at zero or saturation. Second, tasks demanded genuine multi-step planning: each sandbox attempt ran for up to **1.5 hours**, spanning more than **100 tool calls** and over **256K output tokens**. Third, tasks with exploitable grading shortcuts were audited and repaired or discarded. To stretch learning across sessions reaching **1 million tokens** without exceeding the 256K-token native window, RadixArk used harness summarization across chained trajectories. ### Asynchronous RL and the 80% Sandbox Cost Cut Training agents that think and execute shell commands for 90 minutes per rollout normally starves training GPUs while trainers wait on slow environments. RadixArk solved that bottleneck using **Miles** in fully asynchronous mode on **144 NVIDIA H200 GPUs** over 18 days (63,000 GPU-hours total): 128 GPUs continuously generated agent rollouts while 16 GPUs ran weight updates on an internal training backend, pushing fresh weights back out after every step. According to RadixArk's telemetry, that backend achieved **37% model FLOPs utilization (MFU)**, compared with **16% MFU** on a comparable open-source Megatron trainer—a **57% reduction in GPU time** and roughly double the end-to-end throughput (**16,800 tokens/second** versus 8,300). On the environment side, RadixArk built its execution layer on the open-source **AgentENV** distributed sandbox platform, running roughly **944,000 sandboxes** between mid-September and mid-November at up to **1,566 concurrent instances**, dynamically sizing sandbox memory to cut cloud sandbox spending by roughly **80%** and lifting creation reliability above **99.95%** from late October onward. ### Reading the Benchmarks—and Their Caveats Across RadixArk's evaluations (using `temperature=1.0`, `top_p=0.95`, `top_k=20`, and `xhigh` reasoning effort), DaisyRose-27B improves on its `Qwen3.8-27B` base across all seven reported suites: - **Terminal-Bench 4.1:** 12.0% pass@1 (up from 5.6% on Qwen3.8-27B; averaged over 8 repetitions with `mini-swe-agent` summarization) - **Terminal-Bench 2.1:** 75.5% (up from 73.0%) - **SWE-Bench Pro v2:** 94.7% (up from 91.3%) - **DeepSWE v1.1:** 46.2% (up from 42.2%) - **IFBench:** 82.4% (up from 79.5%, beating every larger open-weight model listed in RadixArk's comparison) - **GPQA Diamond:** 90.7% (up from 89.2%, edging out GPT-6 Luna's 90.5%) - **Humanity's Last Exam (text-only):** 34.3% (up from 33.9%) Two methodological details merit scrutiny before treating the scoreboard as an apples-to-apples sweep. First, RadixArk reports the higher of its BF16 and FP8 runs per benchmark (noting the two stayed within one percentage point where both ran). Second, because several reference baselines lacked Terminal-Bench 4.1 entries, the model card substitutes Terminal-Bench 4.0 scores for `Qwen3.8-27B`, `Muse Glimmer-30B`, `Qwen3.6-35B-A3B`, and `Muse Spark Open`, while `Inkling 2 Small` (594B total / 17B active) and `MiMo V2.6 Flash` (309B total / 15B active) still lead on raw coding benchmarks like DeepSWE v1.1. Even with those caveats, matching or outscoring 500B–800B MoE models on IFBench and GPQA Diamond at 27B dense parameters demonstrates how much headroom remains in post-training curation alone. ### Single-GPU Deployment and Hybrid GDN Tuning For practitioners deploying locally, the [SGLang Cookbook](https://docs.sglang.io/cookbook/autoregressive/RadixArk/DaisyRose-27B) lists single-GPU configurations on NVIDIA H200, GB300, RTX PRO 6000, RTX 5090, and DGX Spark across BF16, FP8, and NVFP4 checkpoints. Combining DaisyRose-27B with RadixArk's **DSpark** speculative draft model (`RadixArk/Qwen3.8-27B-DSpark`) yields **3.4 to 3.5 accepted tokens per step** on GB300 in long-context thinking workloads. One non-obvious operational pitfall highlighted in SGLang's documentation involves the `--mamba-full-memory-ratio` flag. Because the Qwen3.8-27B architecture uses hybrid Gated DeltaNet (GDN) layers alongside paged attention, SGLang's default `0.9` ratio over-provisions the KV cache pool and quietly throttles concurrency; calculating the ratio from expected average request length and GDN state slots unlocks full multi-request throughput on a single card. ### Sources & Further Reading - [Hugging Face Blog — DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin) - [RadixArk Official Launch Article — DaisyRose-27B](https://www.radixark.com/blog/daisyrose-27b) - [SGLang Documentation — Deploying DaisyRose-27B With DSpark Speculative Decoding](https://docs.sglang.io/cookbook/autoregressive/RadixArk/DaisyRose-27B) - [GitHub — kvcache-ai/AgentENV Distributed Agent Sandbox](https://github.com/kvcache-ai/AgentENV) - [Hugging Face Model Card — RadixArk/DaisyRose-27B](https://huggingface.co/RadixArk/DaisyRose-27B) ~AI ~OpenSourceAI ~Qwen ~ReinforcementLearning ~SGLang