TL;DR
- RadixArk’s first model, DaisyRose-27B, is built on Qwen3.8-27B using reinforcement learning only. No SFT, no distillation.
- Terminal Bench 4.1 pass@1 rose from 5.6 on the base model to 12.0, after 18 days on 144 H200 GPUs.
- It’s released under Apache-2.0 in BF16, FP8 and NVFP4, and it still trails MiMo V2.6 Flash, GPT-6 Luna and Inkling 2 Small.
- The scores are self-reported, run on an internal mini-swe-agent harness. Treat them as a claim, not a verdict.
RadixArk’s first model skips the usual shortcuts
RadixArk has released DaisyRose-27B, and the model card is blunt about the method. It’s a multimodal reasoning model built on Qwen3.8-27B “with reinforcement learning only, without any supervised fine-tuning or distillation.” The details sit in the Hugging Face repo from Xylolabs, which hosts the GGUF quantized variants.
The training run lasted 18 days on 144 H200 GPUs. The data was about 1,200 curated long-horizon agent tasks, with sessions stretching up to 1M tokens. That’s a small task set for such a long run.
The result: Terminal Bench 4.1 pass@1 climbed from 5.6 for the Qwen3.8-27B base to 12.0. The weights ship under Apache-2.0 in BF16, FP8 and NVFP4, and the 256K-token context window carries over unchanged.
A modest score, a cleaner experiment
Start with the uncomfortable part. A 12.0 on Terminal Bench 4.1 doesn’t win anything. MiMo V2.6 Flash scores 16.1, GPT-6 Luna scores 15.2, and Inkling 2 Small sits at 13.0. DaisyRose-27B is behind all three, and only one point short of the closest.
But the leaderboard position isn’t the interesting bit. The base model scored 5.6. RL alone pushed it past double that, with no demonstration data and no bigger teacher model to copy from. SFT and distillation are the familiar routes to new behaviour. Dropping both and still moving the number this far is a cleaner experiment than most releases bother to run.
Think of it as retuning a car’s engine while the chassis, the keys and the fuel tank stay exactly as they were. Same body, same controls, a very different pull on the road. That’s what makes the result readable: almost nothing else changed, so the gain points at the training method.
I’d argue the small dataset is the more provocative detail than the score. Roughly 1,200 tasks is not a mountain of data. Is that enough to call this a recipe other teams can copy, or did RadixArk pick unusually well-matched tasks? The card doesn’t settle it.
And then there’s the harness. The numbers come from RadixArk’s internal mini-swe-agent setup, not from a third party. Does a 5.6 to 12.0 jump hold up when someone else runs the test? I wouldn’t bet the roadmap on it yet. I’d happily bet an afternoon of GPU time on checking.
The compute is worth a sentence too. 144 GPUs for 18 days works out to 2,592 GPU-days. Real money, real hardware. But the lesson here is about method, not about who owns the biggest cluster.
Drop-in by design
DaisyRose-27B keeps the Qwen3.8-27B architecture, tokenizer, chat template, image understanding and 256K-token context window. In practice that means it’s pitched as a drop-in replacement. If your pipeline already talks to the base model, the swap shouldn’t mean rewriting prompt templates or tokenization.
That matters because agent gains usually arrive with strings attached. Here the pitch is simpler: same interface, better terminal behaviour on at least one benchmark.
The release formats cover a range of hardware. BF16 is the full-precision option, FP8 and NVFP4 are lower-precision variants, and Xylolabs supplies GGUF quantizations for people running models locally. Apache-2.0 means commercial use doesn’t require a negotiation.
Still, the base model remains the reference point. RadixArk is a first-time model builder, and DaisyRose-27B is a fine-tune rather than a ground-up system. The achievement is the training approach. The foundation belongs to Qwen.
Three checks that decide whether this recipe travels
First, independent runs of Terminal Bench 4.1. If outside testers land near 12.0 using their own harness, the claim firms up fast. If they don’t, the internal mini-swe-agent setup becomes the story.
Second, whether RadixArk shares more about the task set and training setup. The 1,200 tasks and 1M-token sessions are the heart of the method, and anyone hoping to reproduce it will want to know what made them work.
Third, whether the approach shows up on other base models. A single result on one base is a data point. A repeat on another is a pattern. Also keep an eye on the gap to Inkling 2 Small at 13.0. One point is close enough that a follow-up release could erase it.
Editor's Note
What caught my eye wasn't the 12.0, which loses to three named rivals. It was how small the task set is. About 1,200 tasks, and the score still more than doubled. If that holds up outside RadixArk's own harness, plenty of teams sitting on open base models should be annoyed they didn't try it first. I'm waiting for someone independent to rerun it before I get excited.
– Sanket Chaukiyal, founder, SmartChunks
FAQ
What is DaisyRose-27B?
It's RadixArk's first model, a multimodal reasoning model built on Qwen3.8-27B. It was trained with reinforcement learning only, without supervised fine-tuning or distillation, and it's released under Apache-2.0.
How much did it improve on Terminal Bench 4.1?
Pass@1 rose from 5.6 for the Qwen3.8-27B base model to 12.0. That's self-reported by RadixArk using an internal mini-swe-agent harness, not independently verified.
Does it beat other models on that benchmark?
No. MiMo V2.6 Flash scores 16.1, GPT-6 Luna 15.2 and Inkling 2 Small 13.0, all above DaisyRose-27B's 12.0. It does clearly beat its own base model at 5.6.
Can it replace Qwen3.8-27B in an existing setup?
It's designed as a drop-in replacement. It keeps the base architecture, tokenizer, chat template, image understanding and 256K-token context window, and ships in BF16, FP8 and NVFP4, with GGUF variants from Xylolabs.
