RadixArk releases DaisyRose-27B, an Apache 2.0 RL post-trained Qwen3.8-27B variant that raises Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
RadixArk published DaisyRose-27B, its first model, as BF16, FP8 and NVFP4 checkpoints under Apache 2.0. It is built on Qwen3.8-27B and trained with RL (Miles framework) on about 1,200 curated long-horizon agent tasks. The run took 18 days on 144 H200 GPUs (about 63,000 GPU-hours), with sessions up to 1M tokens via harness summarization. Architecture, tokenizer, chat template, image understanding and 256K context are unchanged. Reported benchmark gains over the base model are Terminal-Bench 4.1 5.6->12.0, Terminal-Bench 2.1 73.0->75.5, DeepSWE 42.2->46.2, SWE-Bench Pro v2 91.3->94.7, HLE 33.9->34.3, IFBench 79.5->82.4 and GPQA Diamond 89.2->90.7. The article also claims RadixArk's internal training backend reached 37% MFU against 16% for the open-source trainer, a 57% GPU-time reduction. It reports about 80% lower sandbox cloud costs, and 99% sandbox creation success (over 99.95% since late October).
Entities: RadixArk, DaisyRose-27B, Qwen3.8-27B, Qwen team, Miles, Eigent AI
1 primary
What happened
RadixArk, a new company, released DaisyRose-27B, its first model, under Apache 2.0 in BF16, FP8 and NVFP4 versions. It is Qwen3.8-27B further trained with reinforcement learning on about 1,200 curated long-running agent tasks, taking 18 days on 144 H200 GPUs (roughly 63,000 GPU-hours). Architecture, tokenizer, chat template, image understanding and 256K context are unchanged. The biggest reported gain is Terminal-Bench 4.1 pass@1, up from 5.6% to 12.0%. Other gains are small: SWE-Bench Pro v2 91.3 to 94.7, DeepSWE 42.2 to 46.2, Terminal-Bench 2.1 73.0 to 75.5, and HLE 33.9 to 34.3.
Why it matters
Teams building coding and terminal agents get a freely licensed 27B model that can be self-hosted, and because the weights are open, anyone can check the claims. The Terminal-Bench 4.1 result suggests a modest, targeted RL run can lift long-horizon agent work on a mid-sized base. Most other benchmark gains are only a few points and may fall within run-to-run variation. The 12.0% score is still low in absolute terms, so this is not a sign that the model can reliably do real terminal tasks unattended. The efficiency claims matter mainly to labs doing their own RL, and only if RadixArk's backend or method becomes available to them.
What is noise
"Pushing the frontier" overstates a fine-tune of someone else's base model whose headline gain starts from a very low 5.6%. The 37% versus 16% MFU, 57% GPU-time saving, 80% lower sandbox cost and 99% sandbox success figures are all RadixArk's own numbers, from an internal backend, with no independent verification and no stated baseline setup. The extraction also lists no primary links, so the benchmark claims cannot yet be traced to a reproducible evaluation harness or settings.
Watch next
- 01Independent reproduction of the Terminal-Bench 4.1 and SWE-Bench Pro v2 results on the released weights, with the same harness settings and several seeds, within the next few weeks.
- 02Whether RadixArk releases its training backend, task set or sandbox tooling, or only the weights, since the 57% cost and 37% MFU claims cannot be tested without them.
- 03Whether the gains hold on tasks outside the 1,200 training-style tasks, such as fresh agent benchmarks and real developer workflows, and whether quantised FP8 and NVFP4 versions keep the improvement.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680