RadixArk's DaisyRose-27B model, built on Qwen3.8-27B and trained with reinforcement learning using Miles, achieves significant improvements over the original Qwen3.8-27B model on various benchmarks, with a notable 12.0% pass@1 score on Terminal Bench 4.1, up from 5.6%. The model was trained on 1,200 curated tasks in 18 days on 144 H200 GPUs, with an efficient training backend and robust sandbox service. AI summary
Firehose
Filtered to Companies, tagged “transformer models” · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives
Browse by tag
Transformers has implemented expert streaming to offload MoE model weights, achieving speeds of ~14 tok/s, 6% behind Strata's 13.9 tok/s, by streaming only the routed experts and using a bounded cache that groups_mm runs on as is. AI summary