# DaisyRose-27B: Agent RL on Qwen3.8 with Apache 2.0 Weights

> Source: <https://explainx.ai/blog/daisyrose-27b-radixark-agent-rl-qwen-november-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-19 00:00:00+00:00

**Key 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.

**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**.

The 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.

## TL;DR benchmark table

RadixArk published this comparison on the Hugging Face blog:

| 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 | 

DaisyRose-27B benchmark gains on Terminal-Bench and SWE tasks versus Qwen3.8-27B
*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*

Scores 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.

## Training recipe: tasks, sandboxes, and async RL

Three ingredients made the run feasible, per RadixArk:

**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.

**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.

**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.

That 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.

## How to serve DaisyRose-27B

RadixArk 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"`.

The 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.

If 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.

## Open weights: what you get on Hugging Face

Checkpoints ship as:

- **BF16** full precision
- **FP8** and**NVFP4** for NVIDIA-friendly serving

Architecture, tokenizer, **256K context**, and **image understanding** stay aligned with Qwen3.8-27B, lowering migration friction.

For 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.

## Why DaisyRose matters in the agent model lane

Frontier 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.

It also highlights **training systems** competition:

| Layer | DaisyRose story | 
|---|---|
| Data | Eigent-partnered mistake-targeted tasks | 
| RL framework | Miles (open source) | 
| Sandboxes | AgentENV at million-task scale | 
| Serving | SGLang plus DSpark | 

Teams 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.

## Miles and async RL: what the system diagram implies

RadixArk 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**.

The blog contrasts RadixArk’s internal training backend with a **Megatron-based open trainer** on comparable runs:

| Metric (per RadixArk blog) | Internal backend | Open Megatron-class trainer cited | 
|---|---|---|
| Model FLOPs utilization | **37%** peak MFU | **16%** on comparable runs | 
| GPU cost for same work | **57% lower** GPU time claimed | baseline | 
| End-to-end tokens/s on 144 H200s | **~16,800** | **~8,300** | 
| MoE scaling claim | **6.2×** on Qwen3.6-35B-A3B | not independently verified | 
| Large GLM MoE claim | **>10×** on ~300B GLM-5.3-Flash | not independently verified | 

Those are **vendor train-stack benchmarks**, not application-level guarantees. They matter if you operate your own RL farm and pay for GPU-hours directly.

GPU efficiency comparing RadixArk backend with Megatron-class open trainer
*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*

## Million-token sessions and harness summarization

Tasks allowed extremely long agent jobs. The blog mentions sessions up to **1M tokens**, folded into **multiple trajectories** via **harness summarization** so learning remains tractable.

That 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.

If 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.

## Sandbox reliability at RL scale

AgentENV sandbox scale during DaisyRose-27B training
*Image: [RadixArk](https://huggingface.co/blog/RadixArk/daisyrose-27b-pushing-the-frontier-of-post-trainin?ref=explainx)*

AgentENV statistics in the post are a **supply-chain signal** for anyone copying this recipe:

- **~944,000** sandboxes between mid-September and mid-November 2026
- **~1,566** peak concurrent sandboxes
- **99%+** creation success overall,**>99.95%** since late October
- Infrastructure scaling from **one machine to 12** with**ten upgrades** during training

RadixArk 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.

Compare 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.

## Download, eval, and ship checklist

1. Pull **BF16, FP8, or NVFP4** weights from Hugging Face listings linked in the announcement.
2. Stand up **SGLang** using RadixArk’s cookbook paths for your GPU generation.
3. Enable **DSpark speculative decoding** if your hardware matches the documented configs.
4. Run **Terminal-Bench 4.1** with the harness versions named on the model card before you publish internal scorecards.
5. Log failures with repro steps in the Hugging Face **Community** tab so RadixArk can distinguish harness bugs from model bugs.

For 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.

## Research lineage and open-source stack

RadixArk explicitly thanks **Eigent AI** for training data partnership, the **Qwen team** for the base model, and communities around **Miles**, **SGLang**, **slime**, **AgentENV**, and **Harbor**.

If you want to train your own variant:

- **Miles** provides the RL orchestration layer (open source).
- **AgentENV** supplies sandbox isolation (open source with RadixArk contributions).
- **Qwen3.8-27B** supplies tokenizer and multimodal hooks preserved in DaisyRose.

That 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.

## Limitations

Terminal-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.

RadixArk invites failure reports in the model **Community** tab. Treat social hype about "frontier killer" narratives as unverified unless you replicate benchmarks.

SWE-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.

## When to pick DaisyRose over API-only coding models

Choose 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.

Hybrid 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.

Read 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.

For 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.

Document **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.

If 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.
