Summary:

RadixArk’s first model ships in BF16, FP8 and NVFP4. Compare weight sizes and see what its reported coding gains mean for local evaluation.

RadixArk released DaisyRose-27B on November 18. Its first model builds on Qwen3.8-27B with reinforcement learning and is downloadable under Apache 2.0 in three formats. RadixArk reports gains on multi-step coding work, while the released model retains Qwen’s text-and-image inputs and 262,144-token context window.

Which DaisyRose-27B download should you choose?

Official buildApproximate weight-file size
BF1655.6 GB
FP830.9 GB
NVFP421.9 GB

Sizes are sums of the Safetensors files, rounded in decimal GB. They do not include runtime overhead. Before choosing a smaller package, check that your serving stack and hardware support its format and that it leaves enough memory for your expected prompt lengths.

What the benchmark gains show

RadixArk reports 12.0% pass@1 for DaisyRose versus 5.6% for its Qwen base in the team’s Terminal-Bench 4.1 before-and-after comparison. Even after the gain, the absolute result is 12%. Do not compare that number with entries from Terminal-Bench 4.0 or with a run using a different harness.

The following selected rows use the same named benchmark editions in RadixArk’s model card. The card combines RadixArk’s runs with published baseline results, so test configurations may differ. The highest displayed score on each row is bolded.

BenchmarkDaisyRose-27BQwen3.8-27BMuse Glimmer-30B
Terminal-Bench 2.175.5%73.0%51.7%
DeepSWE v1.146.2%42.2%4.4%
SWE-Bench Pro v2 (full)94.7%91.3%80.4%
IFBench82.4%79.5%77.0%

For the full SWE-Bench Pro v2 row, the model card cites the 642-task baseline table and reports a mini-swe-agent setup for DaisyRose. On IFBench it uses the official scorer and loose prompt-level accuracy. The card selects the higher of its BF16 and FP8 results where both exist; it does not establish the same scores for NVFP4.

Why the training and serving limits differ

RadixArk says it used roughly 1,200 curated tasks and trained for 18 days on 144 H200 GPUs. Some training sessions went up to one million tokens. That figure describes training sessions, not the context size of the downloaded checkpoint. The model card lists 262,144 tokens for the release.

If you already use Qwen3.8-27B, start with a small set of repeatable jobs and compare completion rate, time and resource use at the same prompt lengths. That is more informative for a model swap than assuming the benchmark gap will reproduce on your workload.

Sources