以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Gated DeltaNet-2

6 atoms · 跨 1 天 · 首见 2026-05-21 · 最近 2026-05-21

三色: 🟦 fact 0 · 🟥 take 6 · stance ▲3/▼0/◆0

来源: X 6

时态: fresh:6

标签: 好数字:3 · 好观点:2 · 好信源:1

2026-05-21 · 模型发布
Gated DeltaNet-2 is here. 🚀
tags: 好信源 · → daily
2026-05-21 · 性能对比
Gated DeltaNet-2 outperforms KDA and Mamba-3, the latest and best recurrent architectures, head to head at 1.3B.
tags: 好数字 · → daily
2026-05-21 · 评测结果-语言建模与推理
Best average on language modeling + commonsense reasoning, in both recurrent and hybrid settings
tags: 好数字 · → daily
2026-05-21 · 长上下文检索评测结果
Biggest gains on long-context RULER retrieval. S-NIAH-3 jumps from 63 to 90 over KDA, and multi-key needle retrieval climbs from 28 to 38
tags: 好数字 · → daily
2026-05-21 · 技术架构创新
_Gated DeltaNet-2 decouples them. ✂️ a channel-wise erase gate b_t picks which key-side coordinates to read and remove ✍️ a channel-wise write gate w_t picks which value-side coordinates to commit_
tags: 好观点 · → daily
2026-05-21 · 训练效率
⚡ still trains fast: chunkwise WY algorithm with gate-aware backward, fused in Triton
tags: 好观点 · → daily