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

Kimi K2.5

21 atoms · 跨 11 天 · 首见 2026-04-20 · 最近 2026-06-27

三色: 🟦 fact 17 · 🟥 take 4 · stance ▲5/▼1/◆3

来源: X 21

时态: fresh:21

标签: 好数字:15 · 好信源:3 · 好观点:3 · 好思考:1

叙事 (narrative) (5)

2026-05-18 · 被 Cursor 用于 Composer 2.5 构建
it's built on Kimi K2.5
tags: 好信源 · → daily
2026-05-14 · scaling characteristics for pass@k
Why? this isn't about linear TTC scaling ... in pass@k, undertrained models should have more gain with k
tags: 好思考 · → daily
2026-06-17 · 开放模型里程碑
the highest-ranking model on Design Arena is, for the first time ever, an open model
tags: 好观点 · → daily
value: importance=highest ranking open model on Design Arena ever
2026-05-26 · Agent能力
Kimi is highly capable (for more simplistic agent behaviors) in the current agent paradigm
tags: 好观点 · → daily
value: direction=capable for simpler agent behaviors
2026-05-14 · ARC-AGI score
It's not, it's on the public set. Still impressive.
tags: 好观点 · → daily

事实 (fact) (16)

2026-06-17 · Design Arena排名
Kimi K2.5 is tied for #1 on Design Arena, in the same performance band as Gemini 3 and Opus 4.5
tags: 好数字·好信源 · → daily
2026-06-27 · benchmark score
Kimi K2.5 > GLM 5.2 and Opus 4.7
tags: 好数字 · → daily
value: qty=4.5
2026-06-20 · ARC-AGI-2 得分
Kimi K2.5 released on January 27, 2026, with a score of 11.8%
tags: 好数字 · → daily
value: qty=11.8% · date=2026-01-27
2026-06-17 · Design Arena 排名
Kimi K2.5 is tied for #1 on Design Arena, in the same performance band as Gemini 3 and Opus 4.5
tags: 好数字 · → daily
value: rank=tied for #1
2026-05-26 · ARC-AGI-2得分
Kimi K2.5 ARC-AGI-2得分: 10.8
tags: 好数字 · → daily
value: qty=10.8
2026-05-26 · ARC-AGI-2官方榜单得分
11.8% on official leaderboard (V3.2 4.0%)
tags: 好数字 · → daily
value: qty=11.8%
2026-05-26 · 预训练+后训练计算量
cursor says their extended pretrain + post train of Kimi was 7x the baseline for k2.5
tags: 好数字 · → daily
value: qty=7x baseline · direction=extended pretrain + post train
2026-05-26 · 基线训练量
baseline for k2.5, which is 30T tokens and 30B active params
tags: 好数字 · → daily
value: qty=30T tokens 30B active params
2026-05-26 · 训练算力需求
K2.5 2x that
tags: 好数字 · → daily
value: qty=2倍 K2
2026-05-14 · official semi-private leaderboard score (pass@2)
Official semi-private leaderboard is pass@2, there Kimi K2.5 gets 11.8% with DSV3.2 4.0%.
tags: 好数字 · → daily
value: qty=11.8%
2026-05-11 · performance gap to K2.6
the gap is literally 5 times smaller than between Kimi K2.5 to K2.6 (though identical in secondary metrics)
tags: 好信源 · → daily
2026-05-06 · 最终资金量
Kimi K2.5 最终资金量: 约 £10,000
tags: 好数字 · → daily
value: qty=£10,000 · date=2026-05-06
2026-04-26 · training tokens
K2.5 was trained on another 15T tokens
tags: 好数字 · → daily
value: qty=15T · date=na
2026-04-25 · 推理吞吐量
Kimi K2.5 (8x RTX Pro 6000): 74.39 tok/s
tags: 好数字 · → daily
value: qty=74.39 tok/s · context=8x RTX Pro 6000
2026-04-20 · 幻觉率
Kimi K2.5 scores 6 on the AA-Omniscience Index, primarily driven by low hallucination rate
tags: 好数字 · → daily
2026-04-20 · AA-Omniscience得分
Kimi K2.5 scores 6 on the AA-Omniscience Index, primarily driven by low hallucination rate
tags: 好数字 · → daily
value: qty=6

时间轴 (近 20)

2026-06-27 · fact · benchmark score · → daily
2026-06-20 · fact · ARC-AGI-2 得分 · → daily
2026-06-17 · fact · Design Arena 排名 · → daily
2026-06-17 · narrative · 开放模型里程碑 · → daily
2026-06-17 · fact · Design Arena排名 · → daily
2026-05-26 · fact · ARC-AGI-2得分 · → daily
2026-05-26 · fact · ARC-AGI-2官方榜单得分 · → daily
2026-05-26 · fact · 预训练+后训练计算量 · → daily
2026-05-26 · fact · 基线训练量 · → daily
2026-05-26 · narrative · Agent能力 · → daily
2026-05-26 · fact · 训练算力需求 · → daily
2026-05-18 · narrative · 被 Cursor 用于 Composer 2.5 构建 · → daily
2026-05-14 · narrative · ARC-AGI score · → daily
2026-05-14 · fact · official semi-private leaderboard score (pass@2) · → daily
2026-05-14 · narrative · scaling characteristics for pass@k · → daily
2026-05-11 · fact · performance gap to K2.6 · → daily
2026-05-06 · fact · 最终资金量 · → daily
2026-04-26 · fact · training tokens · → daily
2026-04-25 · fact · 推理吞吐量 · → daily
2026-04-20 · fact · 幻觉率 · → daily

← 实体目录 · 系统日志