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
MiniMax M2
7 atoms · 跨 1 天 · 首见 2026-05-27 · 最近 2026-05-27
三色: 🟦 fact 7 · 🟥 take 0 · stance ▲2/▼0/◆0
来源: X 7
时态: fresh:7
标签: 好观点:4 · 好数字:2 · 好信源:1
事实 (fact) (7)
- 2026-05-27 · 尝试混合滑动窗口注意力变体
They tried hybrid sliding-window attention variants (like so many others, like Xiaomi MiMo, Laguna, Gemma 4, Arcee, Olmo 3, etc.)
tags: 好信源 · → daily
- 2026-05-27 · 细粒度MoE相比基线提升MATH分数
The fine-grained setup improves MATH from 19.6 to 24.1
tags: 好数字 · → daily
value: qty=从19.6到24.1
- 2026-05-27 · 细粒度MoE相比基线提升HumanEval分数
HumanEval from 29.7 to 32.5
tags: 好数字 · → daily
value: qty=从29.7到32.5
- 2026-05-27 · 混合滑动窗口注意力生产质量权衡不值得
they said that the production-quality tradeoffs were not worth it for M2
tags: 好观点 · → daily
- 2026-05-27 · 线性和稀疏注意力在低精度存储KV状态时更脆弱
they found that these efficient attention variants may be more fragile when KV-like state or intermediate memory is stored in lower precision
tags: 好观点 · → daily
- 2026-05-27 · 线性和稀疏注意力前缀缓存支持更差
they have worse prefix caching support, which matters a lot when using coding agents (which reuse a lot of the context)
tags: 好观点 · → daily
- 2026-05-27 · 训练用于软件工程任务的类Agent行为现在是训练管线重要组成部分
this papers confirms that training for agent-like behavior on software engineering task is now a big component of the training pipeline.
tags: 好观点 · → daily
时间轴 (近 20)
- 2026-05-27 ·
fact · 尝试混合滑动窗口注意力变体 · → daily
- 2026-05-27 ·
fact · 混合滑动窗口注意力生产质量权衡不值得 · → daily
- 2026-05-27 ·
fact · 线性和稀疏注意力在低精度存储KV状态时更脆弱 · → daily
- 2026-05-27 ·
fact · 线性和稀疏注意力前缀缓存支持更差 · → daily
- 2026-05-27 ·
fact · 细粒度MoE相比基线提升MATH分数 · → daily
- 2026-05-27 ·
fact · 细粒度MoE相比基线提升HumanEval分数 · → daily
- 2026-05-27 ·
fact · 训练用于软件工程任务的类Agent行为现在是训练管线重要组成部分 · → daily
← 实体目录 · 系统日志