{"slug": "daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights", "title": "DaisyRose-27B: Agent RL on Qwen3.8 with Apache 2.0 Weights", "summary": "RadixArk released DaisyRose-27B on November 18, 2026, its first open model, built by post-training Qwen3.8-27B with reinforcement learning on roughly 1,200 long-horizon agent tasks and published under Apache 2.0 in BF16, FP8 and NVFP4 weights. RadixArk reported Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0% versus the base Qwen3.8-27B checkpoint, with smaller gains on DeepSWE v1.1 (42.2 to 46.2) and SWE Bench Pro v2 (91.3 to 94.7). RadixArk said the run used 128 H200s for inference rollouts and 16 for training, about 944,000 AgentENV sandboxes between mid-September and mid-November 2026, and claimed 57% lower training cost on MFU metrics.", "body_md": "**RadixArk released DaisyRose-27B on November 18, 2026** as its first open model: **Qwen3.8-27B** post-trained with **reinforcement learning** on roughly **1,200 long-horizon agent tasks**. The [Hugging Face announcement](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx) and [model card](https://huggingface.co/RadixArk/DaisyRose-27B?ref=explainx) publish **Apache 2.0** weights in **BF16, FP8, and NVFP4**.\n\nThe headline metric for builders is **Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0%** versus the base Qwen3.8-27B checkpoint. That is still a hard benchmark, but the delta shows what **task-specific RL** buys when sandboxes allow **hundreds of tool calls** and **256K+ output tokens** per attempt.\n\n## TL;DR benchmark table\n\nRadixArk published this comparison on the Hugging Face blog:\n\n| Benchmark | Qwen3.8-27B | DaisyRose-27B | \n|---|---|---|\n| Terminal Bench 4.1 | 5.6 | **12.0** | \n| Terminal Bench 2.1 | 73.0 | 75.5 | \n| DeepSWE v1.1 | 42.2 | 46.2 | \n| SWE Bench Pro v2 | 91.3 | 94.7 | \n| HLE (text-only) | 33.9 | 34.3 | \n| IFBench | 79.5 | 82.4 | \n| GPQA Diamond | 89.2 | 90.7 | \n\nDaisyRose-27B benchmark gains on Terminal-Bench and SWE tasks versus Qwen3.8-27B\n*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*\n\nScores use the higher of BF16 and FP8 runs where both were evaluated; the blog says they differ by less than one point in those cases.\n\n## Training recipe: tasks, sandboxes, and async RL\n\nThree ingredients made the run feasible, per RadixArk:\n\n**Curated tasks with Eigent AI.** Thousands of candidates collapsed to **about 1,200** kept tasks. Selection rules included targeting **concrete base-model mistakes**, requiring **multi-step dependent work** (up to **1.5 hours** per sandbox attempt, **100+ tool calls**, **256K+ output tokens**), and ensuring rewards require **actually solving** the intended problem.\n\n**Miles RL in async mode.** **128 H200s** ran inference rollouts while **16** trained the policy, pushing updated weights back after each step. RadixArk claims **~16,800 tokens/s** end-to-end across **144 GPUs** versus **~8,300** on a Megatron-class open trainer on comparable runs, and **57% lower training cost** on MFU metrics for this release.\n\n**AgentENV sandboxes.** About **944,000 sandboxes** between mid-September and mid-November 2026, peaking around **1,566 concurrent**, with **99%+ creation success** (above **99.95%** since late October per the post). RadixArk says cloud costs were about **80% lower** than popular hosted sandbox vendors, partly via elastic memory allocation.\n\nThat infrastructure story matters if you are comparing [Prime Intellect sandboxes](https://explainx.ai/blog/prime-sandboxes-filesystem-checkpoints-fork-rollouts-agentic-rl-guide-2026) or [MIT self-improvement harness research](https://explainx.ai/blog/llm-agents-self-improve-without-prescribed-loop-mit-isola-2026) for your own RL loop.\n\n## How to serve DaisyRose-27B\n\nRadixArk recommends **SGLang** with **DSpark speculative decoding**. After launch with served name `DaisyRose-27B` on port `30000`, use the OpenAI Python client against `http://localhost:30000/v1` with `api_key=\"none\"`.\n\nThe model **thinks by default**. Responses split **`reasoning_content`** and final **` content`**, matching Qwen-style reasoning templates. Temperature **1.0**, top_p **0.95**, and extra_body `top_k: 20` follow the model card quickstart cited on Hugging Face.\n\nIf you run [Qwen3.5 adapters elsewhere](https://explainx.ai/blog/qwen38-27b-opencode-harness-adapter-lora-november-2026), confirm harness compatibility before you swap weights in production.\n\n## Open weights: what you get on Hugging Face\n\nCheckpoints ship as:\n\n- **BF16** full precision\n- **FP8** and**NVFP4** for NVIDIA-friendly serving\n\nArchitecture, tokenizer, **256K context**, and **image understanding** stay aligned with Qwen3.8-27B, lowering migration friction.\n\nFor evaluators, reproduce Terminal-Bench with the **same harness versions** RadixArk cites in the model card. Community scores on different harness configs will not match the table above.\n\n## Why DaisyRose matters in the agent model lane\n\nFrontier labs chase **coding agents** with proprietary RL stacks. DaisyRose is evidence that **mid-size open bases** can gain **agentic terminal skills** when training data enforces long jobs and honest graders.\n\nIt also highlights **training systems** competition:\n\n| Layer | DaisyRose story | \n|---|---|\n| Data | Eigent-partnered mistake-targeted tasks | \n| RL framework | Miles (open source) | \n| Sandboxes | AgentENV at million-task scale | \n| Serving | SGLang plus DSpark | \n\nTeams betting on [higher SWE-Bench merge rates](https://explainx.ai/blog/higher-scoring-coding-models-swe-bench-merge-rate-natural-experiment-2026) should still run **their** repos through DaisyRose before trusting leaderboard deltas.\n\n## Miles and async RL: what the system diagram implies\n\nRadixArk trained with **Miles**, its open reinforcement learning framework, in **fully asynchronous** mode: agent environments keep producing trajectories while the trainer consumes them without lockstep barriers. That design targets agentic workloads where **environment time** dominates **backward pass time**.\n\nThe blog contrasts RadixArk’s internal training backend with a **Megatron-based open trainer** on comparable runs:\n\n| Metric (per RadixArk blog) | Internal backend | Open Megatron-class trainer cited | \n|---|---|---|\n| Model FLOPs utilization | **37%** peak MFU | **16%** on comparable runs | \n| GPU cost for same work | **57% lower** GPU time claimed | baseline | \n| End-to-end tokens/s on 144 H200s | **~16,800** | **~8,300** | \n| MoE scaling claim | **6.2×** on Qwen3.6-35B-A3B | not independently verified | \n| Large GLM MoE claim | **>10×** on ~300B GLM-5.3-Flash | not independently verified | \n\nThose are **vendor train-stack benchmarks**, not application-level guarantees. They matter if you operate your own RL farm and pay for GPU-hours directly.\n\nGPU efficiency comparing RadixArk backend with Megatron-class open trainer\n*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*\n\n## Million-token sessions and harness summarization\n\nTasks allowed extremely long agent jobs. The blog mentions sessions up to **1M tokens**, folded into **multiple trajectories** via **harness summarization** so learning remains tractable.\n\nThat pattern aligns with research explainx.ai tracks on **[latent reasoning monitorability](https://explainx.ai/blog/gpt-6-2-astra-epoch-12-3-latent-reasoning-steps-monitorability-2026)** and **long-context agent harnesses**: the model is not necessarily generating one monologue; the harness compresses prior segments so RL credit assignment still works.\n\nIf you fine-tune agents on [OpenCode-style harnesses](https://explainx.ai/blog/opencode-2-0-code-mode-background-agents-shared-server-sdk-upgrade-guide-2026), study how DaisyRose’s **SGLang plus DSpark** serving path exposes `reasoning_content` separately from answers. Downstream apps must not strip reasoning traces when they replay user-visible chats.\n\n## Sandbox reliability at RL scale\n\nAgentENV sandbox scale during DaisyRose-27B training\n*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*\n\nAgentENV statistics in the post are a **supply-chain signal** for anyone copying this recipe:\n\n- **~944,000** sandboxes between mid-September and mid-November 2026\n- **~1,566** peak concurrent sandboxes\n- **99%+** creation success overall,**>99.95%** since late October\n- Infrastructure scaling from **one machine to 12** with**ten upgrades** during training\n\nRadixArk contributes fixes upstream to **AgentENV** and cites **Harbor** support in the ecosystem thank-you list. If your RL loop stalls on sandbox flakes, DaisyRose’s run is a useful reference for what “good enough” reliability looks like at **six nines-ish** creation rates.\n\nCompare with hosted options in [agent sandbox guides](https://explainx.ai/blog/agent-sandboxes-coding-agents-compared-e2b-modal-daytona-guide-2026) when you choose between building AgentENV yourself and renting sandboxes.\n\n## Download, eval, and ship checklist\n\n1. Pull **BF16, FP8, or NVFP4** weights from Hugging Face listings linked in the announcement.\n2. Stand up **SGLang** using RadixArk’s cookbook paths for your GPU generation.\n3. Enable **DSpark speculative decoding** if your hardware matches the documented configs.\n4. Run **Terminal-Bench 4.1** with the harness versions named on the model card before you publish internal scorecards.\n5. Log failures with repro steps in the Hugging Face **Community** tab so RadixArk can distinguish harness bugs from model bugs.\n\nFor teams comparing against [Microsoft MAI Code 2](https://explainx.ai/blog/microsoft-ignite-2026-complete-coverage-every-announcement-november-2026) or [OpenAI Codex profiles](https://explainx.ai/blog/openai-inside-codex-sandbox-permission-profiles-november-2026), DaisyRose is an **open-weight alternative** you can host, not a drop-in API with the same safety envelopes.\n\n## Research lineage and open-source stack\n\nRadixArk explicitly thanks **Eigent AI** for training data partnership, the **Qwen team** for the base model, and communities around **Miles**, **SGLang**, **slime**, **AgentENV**, and **Harbor**.\n\nIf you want to train your own variant:\n\n- **Miles** provides the RL orchestration layer (open source).\n- **AgentENV** supplies sandbox isolation (open source with RadixArk contributions).\n- **Qwen3.8-27B** supplies tokenizer and multimodal hooks preserved in DaisyRose.\n\nThat is a full stack story similar in spirit to [Prime Intellect’s Model Factory loop](https://explainx.ai/blog/prime-intellect-model-factory-train-deploy-loop-november-2026), but DaisyRose is a **finished checkpoint** you can download today.\n\n## Limitations\n\nTerminal-Bench 4.1 at **12% pass@1** is improvement, not solved. DaisyRose can still fail long-horizon plans, mis-handle novel tools, or drift on out-of-distribution tasks not represented in the 1,200 training set.\n\nRadixArk invites failure reports in the model **Community** tab. Treat social hype about \"frontier killer\" narratives as unverified unless you replicate benchmarks.\n\nSWE-Bench Pro v2 gains (91.3 to 94.7) are incremental on an already strong base. The interesting signal is **agentic terminal** lift where Qwen3.8 was weakest, not universal dominance across every coding metric.\n\n## When to pick DaisyRose over API-only coding models\n\nChoose DaisyRose when you need **self-hosted weights**, **Apache 2.0** terms, and **terminal-heavy agents** without routing customer code through a proprietary API. Stick with hosted frontier APIs when you need vendor-managed safety filters, enterprise DPAs, or turnkey IDE integrations.\n\nHybrid teams sometimes route planning to a frontier model and execution to DaisyRose on private GPUs. Watch **token costs** on 256K contexts and instrument sandbox failures the way [OpenAI’s misalignment harness post](https://explainx.ai/blog/openai-rl-agent-email-forwarding-misalignment-report-november-2026) recommends for long-horizon agents.\n\nRead the Hugging Face **model card** for inference hyperparameters and benchmark provenance before you cite DaisyRose in customer-facing security questionnaires. Open weights shift compliance burden to your deployment: you own patching, logging, and abuse monitoring on the inference host the way you would for any other self-hosted [Qwen-family](https://explainx.ai/blog/qwen38-27b-opencode-harness-adapter-lora-november-2026) derivative.\n\nFor capacity planning, RadixArk’s **144 H200** footprint is a useful order-of-magnitude anchor when you negotiate cluster rentals for your own Miles runs, even if your task count is smaller than 1,200.\n\nDocument **NVFP4** versus **BF16** quality in your own eval suite before you pick a production quantization. RadixArk reports sub-one-point benchmark gaps between formats on several tasks, but your prompts may behave differently on tool-calling edges.\n\nIf you also depend on **image understanding** inherited from Qwen3.8, regression-test vision-heavy agent tasks separately from Terminal-Bench text shells today so multimodal regressions do not hide behind averaged coding scores on leaderboards.", "url": "https://wpnews.pro/news/daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights", "canonical_source": "https://explainx.ai/blog/daisyrose-27b-radixark-agent-rl-qwen-november-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog", "published_at": "2026-11-19 00:00:00+00:00", "updated_at": "2026-11-19 04:43:34.976354+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-research", "ai-tools"], "entities": ["RadixArk", "DaisyRose-27B", "Qwen3.8-27B", "Hugging Face", "Terminal-Bench 4.1", "Eigent AI", "Miles RL", "AgentENV"], "also_reported_by": [], "takeaway": "DaisyRose-27B's RL post-training lifted Terminal-Bench 4.1 pass@1 from 5.6% to 12.0% over Qwen3.8-27B, and the Apache 2.0 weights ship in BF16, FP8 and NVFP4 with a 256K context for SGLang serving.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights", "markdown": "https://wpnews.pro/news/daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights.md", "text": "https://wpnews.pro/news/daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights.txt", "jsonld": "https://wpnews.pro/news/daisyrose-27b-agent-rl-on-qwen3-8-with-apache-2-0-weights.jsonld"}}