Dek: RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.
What DaisyRose-27B is: Qwen3.8-27B with a longer training budget
The name suggests a new model. It's a fine-tune, and RadixArk is upfront about that.
DaisyRose-27B keeps the architecture, tokenizer, chat template, image understanding and 256K-token context of Qwen3.8-27B. Nothing was rebuilt. The training was reinforcement learning only, on long-horizon agent tasks with sessions running up to 1M tokens, and the checkpoints ship under Apache 2.0 in BF16, FP8 and NVFP4.
That's the entire premise: same model, better behaviour on jobs that take hundreds of steps. If you already have Qwen3.8-27B deployed, the comparison is a drop-in replacement rather than a migration.
The Hugging Face post is dated 18 November. Press coverage landed on 20 November.
Terminal-Bench 4.1 doubled and stayed low
Terminal-Bench 4.1 measures whether an agent can complete real work in a command-line environment, and pass@1 asks for a clean first attempt.
Qwen3.8-27B scores 5.6% on it, and DaisyRose-27B scores 12.0%.
Both readings matter. The score more than doubled, which is a large relative gain from a training run that only touched post-training. The absolute number's also low enough that most people would call it a research result. One in eight attempts on a hard terminal task isn't a product you hand to an intern.
The rest of the RadixArk benchmark row
The pattern holds across the table, which is the part that surprised me. Every benchmark improved, and the gains are spread rather than concentrated in one place.
Benchmark | Qwen3.8-27B | DaisyRose-27B |
|---|---|---|
Terminal Bench 2.1 | 73.0 | 75.5 |
DeepSWE v1.1 | 42.2 | 46.2 |
SWE Bench Pro v2 | 91.3 | 94.7 |
IFBench | 79.5 | 82.4 |
GPQA Diamond | 89.2 | 90.7 |
HLE, text only | 33.9 | 34.3 |
Scores use the higher of the BF16 and FP8 runs for each benchmark, and RadixArk says the two formats differ by less than a point wherever both were evaluated. That's a useful detail, because it means quantisation isn't quietly buying part of the gain.
The spread is what you'd expect from agent training. DeepSWE, which tests multi-step software work, moves 4 points, and IFBench, which checks instruction-following, moves nearly 3. GPQA Diamond, a graduate-level science set, barely moves at 1.5.
Knowledge didn't change much. Behaviour did.
How 1,200 tasks taught DaisyRose-27B to finish long jobs
The task set is the interesting engineering, because 1,200 isn't a large number for a training run.
RadixArk built the tasks with Eigent AI from a pool of several thousand candidates, and about 1,200 survived review. Each one targets a mistake the base model actually makes, needs several interdependent steps to complete, and was checked so it couldn't be passed without solving the intended problem.
Each attempt got up to 1.5 hours in a sandbox, with more than 100 tool calls and more than 256K output tokens. That's the shape of a real agent job: read, edit, run, read the error, edit again.
Training on tasks like that's expensive per sample, which is why the usual approach is to use millions of cheap short examples instead. RadixArk went the other way, and most of the release is about making that affordable.
Miles, 37% MFU and the 57% GPU saving
Miles is RadixArk's open-source RL framework, and the release runs it fully asynchronously.
In practice that means 128 H200s generate rollouts on agent tasks while 16 more run the trainer, and the updated model is pushed back to the inference fleet after each step.
Nobody waits for anybody.
The utilisation numbers are the sales pitch: 37% model FLOPs utilisation against 16% for the open-source Megatron-based trainer on comparable runs. RadixArk says that works out to 57% less GPU time for the same training. End to end, the run processed about 16,800 tokens per second across 144 GPUs, against about 8,300 for the open-source trainer on the same hardware.
The whole job took 18 days and roughly 63,000 GPU-hours, and the company says throughput didn't degrade over nine weeks of continuous work. It also reports a 6.2x gain on Qwen3.6-35B-A3B. On the roughly 300B-parameter GLM-5.3-Flash the gain is above 10x, so the framework isn't tuned to one model size.
Treat those numbers as vendor comparisons. They're measured against a specific open-source trainer on comparable runs, not against every training stack in existence.
The sandbox service under DaisyRose-27B
Long agent jobs need somewhere to run, and RadixArk built that too. The service sits on the open-source AgentENV project.
Between mid-September and mid-November it ran about 944,000 sandboxes, with up to 1,566 concurrent. Creation succeeded 99% of the time across nine weeks and above 99.95% since late October, at cloud costs about 80% below popular sandbox providers.
An 80% cost cut on sandboxes sounds like a footnote. Do the arithmetic on 100-plus tool calls per attempt and it stops being one. Sandbox time is one of the reasons agent RL stays expensive, and it's the part most teams underbudget.
What the DaisyRose-27B numbers don't prove
All of these results come from RadixArk, on a base model RadixArk also published against. MarkTechPost repeated the table. Nobody independent has reproduced it.
The absolute agent scores are low, so the honest claim is about cost per point rather than about a new frontier. A team with a smaller GPU budget can now buy this kind of improvement, which is a different and more useful claim than "open models caught up."
There's no hosted price to compare, either. The weights are free under Apache 2.0, and inference costs whatever your hardware costs.
The model card is also explicit that this is a coding agent, not a chat assistant. RadixArk thanks the Qwen team for the base model and says nothing about general conversation quality, which is a hint about where the training budget went.
How to serve DaisyRose-27B
RadixArk recommends SGLang with DSpark speculative decoding. The model returns its reasoning in reasoning_content and its final answer in content, so anything that parses chat completions needs to handle both fields.
What does that mean if you're already running Qwen3.8-27B? Start with the short-horizon tasks in your own stack before pointing it at a multi-hour job. The benchmark story is about finishing long agent work, and that's also where a 12.0% pass rate will show up first.
The interesting question now isn't whether a 27B model can be improved with RL. It's whether asynchronous training plus a cheap sandbox becomes the default way small teams do it. RadixArk published the framework and the numbers, so that argument is now testable by people who don't work there.






