RadixArk lança DaisyRose-27B, um modelo de codificação agente de 27B treinado apenas com RL
DaisyRose-27B duplica a pontuação Terminal Bench 4.1 pass@1, roda em uma única GPU e está disponível sob Apache 2.0.
RadixArk announced DaisyRose-27B, its first open-weight model built on the Qwen3.8-27B architecture and post‑trained solely with reinforcement learning.
The model contains 27.8 B dense parameters, a 262,144‑token context and supports text and image input. It can run on one GPU using 52 GB in BF16, 29 GB in FP8 or 23 GB in NVFP4 on Blackwell GPUs.
Benchmark results show Terminal Bench 4.1 pass@1 rising from 5.6 to 12.0, a more than two‑fold increase, while other scores improve modestly, e.g., DeepSWE +4.0 and IFBench +2.9.
Training used RadixArk's Miles RL framework on 144 H200 GPUs for 18 days, consuming about 63,000 GPU‑hours and processing roughly 944,000 sandboxes.
Weights are hosted on Hugging Face in multiple formats; deployment recommendations include SGLang with DSpark speculative decoding or vLLM.
O que isso significa para você
DaisyRose-27B offers a high‑performing, open‑weight alternative for developers who need agentic coding capabilities without proprietary licenses. Its single‑GPU footprint makes it attractive for local or on‑premise setups, complementing the free chat models that dominate community rankings.
No FreeLLMArena
From our rankings on 21 Nov 2026, based on Arena's blind votes:
- Qwen3.8-27B: #25 among free chat models
Fontes
- RadixArk Ships DaisyRose-27B, a 27B Agentic Coding Model Trained With Reinforcement Learning Only, MarkTechPost, 20 Nov 2026
- RadixArk/DaisyRose-27B · Hugging Face, huggingface.co (official), 20 Nov 2026
Escrito automaticamente a partir das fontes acima, depois verificado pelo nosso código: todo número e nome deve aparecer nelas, e nenhuma frase pode copiar uma fonte. Encontrou um erro? Reporte‑o.
Qual IA você pode usar grátis hoje?
Todo modelo de IA gratuito, classificado por votos humanos cegos, com a memória e velocidade que ele precisa na sua própria GPU.