The notable bet here is that a relatively small, carefully curated set of difficult, extended tasks can improve an existing model’s ability to plan and execute—not just its ability to answer isolated questions. DaisyRose applies reinforcement learning to Qwen3.8-27B without supervised fine-tuning or distillation, targeting the multi-step failures that matter in coding and tool-using work.
Key Capabilities
- Long-horizon agent work: Training tasks chained interdependent problems and allowed extended sandbox sessions, with more than 100 tool calls and up to 1M-token sessions. This targets sustained execution rather than short benchmark-style exchanges.
- Keeps the base model’s multimodal interface: It retains Qwen3.8-27B’s image understanding, tokenizer, chat template, and 262,144-token context. Existing stacks built for that base model can adopt it without changing the interaction format.
- Measured gains, with important caveats: RadixArk reports Terminal Bench 4.1 pass@1 improving from 5.6 to 12.0 and IFBench rising from 79.5 to 82.4 versus the base. Results are developer-reported, and DaisyRose’s scores are the higher of its BF16 and FP8 runs; it does not lead every comparison benchmark.
- Multiple weight formats: BF16, FP8, and NVFP4 releases offer different memory tradeoffs. The BF16 weights require about 52 GB, while NVFP4 requires a Blackwell GPU.
Who It Suits—and the Tradeoff
A good fit for researchers and developers testing open-weight reasoning or coding agents who already use the Qwen3.8 interface and can run a model of this size. Look elsewhere if you need a lightweight model, independently verified benchmark results, or consistently top scores across agent evaluations: the reported improvements are meaningful but benchmark-dependent, and the training recipe still relies on substantial compute.