RadixArk releases DaisyRose-27B with company-reported Terminal-Bench gains
November 18: Hacker News adds that Eigent AI built the tasks and about 944,000 AgentENV sandboxes were created; the card says scores take the higher of BF16 and FP8.
RadixArk’s first model, DaisyRose-27B, is on Hugging Face with company-reported Terminal-Bench 4.1 gains over Qwen3.8-27B.
RadixArk released DaisyRose-27B, its first model, as Apache-2.0 weights in BF16, FP8 and NVFP4. It is a multimodal reasoning model built on Qwen3.8-27B and trained only with reinforcement learning through RadixArk’s Miles framework, without supervised fine-tuning or distillation. The company says training used about 1,200 long-horizon agent tasks, built with Eigent AI, over 18 days on 144 H200 GPUs, totaling 63,000 GPU-hours, with sessions up to 1 million tokens and about 944,000 AgentENV sandboxes. The weights keep the base architecture, tokenizer, chat template, image understanding and 256K context. RadixArk reports Terminal-Bench 4.1 pass@1 rising from 5.6 to 12.0 and IFBench at 82.4 versus 79.5; published scores take the higher of BF16 and FP8 and are the company’s own figures.
- Apache-2.0 BF16, FP8 and NVFP4 checkpoints are on Hugging Face.
- Training used about 1,200 tasks and 63,000 GPU-hours on 144 H200s, RadixArk says.
- Sessions ran up to 1 million tokens; about 944,000 AgentENV sandboxes were created.
- Terminal-Bench 4.1 pass@1 is reported at 12.0 versus 5.6 for the base model.
- IFBench is 82.4 versus 79.5; scores take the higher of BF16 and FP8.
What each outlet adds
beyond the shared facts- Hugging Face trending modelsIFBench scores and that published numbers take the higher of BF16 and FP8.
- Hacker NewsEigent AI tasks, 63,000 GPU-hours, and about 944,000 AgentENV sandboxes.
- Organizations
- RadixArkharness-hubEigent AI
Coverage
oldest first · one row per article · UTCharness-hub releases a Qwen3.8-27B LoRA for the pi agent
Authors report Terminal-Bench 2.1 at 55.8% versus 39.5% for the base model, with 62% fewer tool calls.
LoRA adapter tunes Qwen3.8-27B for the OpenCode coding agent
The publisher reports 51.0% on SWE-bench Pro hard 51 with thinking off, above the base model’s high-effort score.
RadixArk releases DaisyRose-27B, an RL-only Qwen3.8 model
Apache-2.0 BF16, FP8 and NVFP4 weights keep the 256K context; Terminal Bench 4.1 pass@1 moves from 5.6 to 12.0.
RadixArk releases DaisyRose-27B, trained from Qwen3.8-27B
RadixArk says 1,200 reinforcement-learning tasks raised Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%.