AITOPAI热报
ModelsOfficial83°

Moonshot AI 发布 Kimi K3：2.8万亿参数开源MoE模型，支持1M上下文

Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context

Primary sourcemarktechpost@Asif Razzaq
Blog/Media· 11 sources
Read the original
Why we picked it

Moonshot AI 把 2.8 万亿参数的 Kimi K3 开源了，用了新注意力机制，还有 1M 上下文，直接看技术细节就对了。

2026年7月16日，Moonshot AI 发布了 Kimi K3 模型，参数规模达2.8万亿。

Image · marktechpost
Body

该模型采用 MoE 架构，包含896个专家，每次推理激活其中16个。Kimi K3 引入了 Kimi Delta Attention 和 Attention Residuals 机制，支持100万token的上下文窗口。

模型以开源形式发布。

Original · marktechpost

Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context

Moonshot AI released Kimi K3 on July 16, 2026. It is a 2.8-trillion-parameter open MoE model built on Kimi Delta Attention and Attention Residuals, activating 16 of 896 experts. The post Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context appeared first on MarkTechPost .

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Kimi K3, Moonshot AI, MoE, 开源模型.

Primary source verification: published directly by marktechpost with first-party authority. Cross-source corroboration: 11 independent sources have covered this development (including lmarena.ai, elvis, Aravind Srinivas), confirming industry-wide signal strength.

#Kimi K3#Moonshot AI#MoE#开源模型#注意力机制

Keep reading

Recent stories with the same tag or section
How-toBuilders15:35
Chinese content网友用停产傲腾内存无显卡跑通 Kimi K3，每秒输出 1.5 词元

这个折腾太硬核了：一块显卡都没有，靠停产的傲腾内存加至强 CPU 跑通月之暗面那个 2.8 万亿参数的 Kimi K3，用 llama.cpp 推理，1 比特量化最快每秒 1.5 个词元。速度感人但证明了家用硬件也能啃超大模型，对喜欢搞本地部署的玩家很有参考价值。

#Kimi K3#llama.cpp#Intel Optane#月之暗面
Event
IT之家Source
PapersBuildersMulti-source10:59
Chinese contentEvoX 论文提出把 Agent Harness 视作运行时策略

一篇关于 agent 工程的硬核论文：EvoX 用 Rust 写的运行时层去管住模型落地时的工具调用、中断恢复和结果校验，不靠记忆和插件就拿到 691 组测试里 56.2% 的成功率。做 agent 框架或工程化落地的开发者建议看看它对 GLM-5.3 和 Kimi K3 提升最大的原因。

#EvoX#agent#论文#GLM-5.3
Event
arXiv cs.AI@Haoyang Zhang et al. (6)Source
PapersUS vs ChinaOfficial10:35
Chinese content姚期智与杨植麟联合发布 TCSExam 基准，Kimi K3 刷新图论纪录

姚期智和杨植麟一起做了个理论计算机科学基准，用 Lean 4 自动判分，Kimi K3 还顺手刷了个 124 顶点 630 边的图论纪录。做形式化验证或模型评测的人可以看看它的出题思路。

#TCSExam#Kimi K3#Lean 4#推理模型
Event
Primary sourcepandaily@[email protected] (Pandaily)Source
PapersBuildersMulti-sourceSelected10:28
Together Compiler Hits 2.6x Speedups on Frontier Model Inference

An agent loop that continuously recompiles inference engines against live traffic could shift how serving providers keep pace with rapidly changing model architectures, replacing static hand-tuned kernels. Teams running inference infrastructure or building MoE serving stacks can study the tiered persistent engine design and context-management ablations for their own optimizer setups. The results also benchmark directly against open source engines on Chinese frontier models like DeepSeek v4.1 Flash and Kimi K3, a useful comparison point when evaluating serving providers.

#Together AI#DeepSeek#Kimi#inference optimization
Event
arXiv: DeepSeek@Reyna Abhyankar et al. (12)Source
ProductsUS vs ChinaMulti-sourceSelected01:12
Together Compiler Hits 2.7× Throughput Gain on B200 and B300

The gap between 1.5× and 2.7× across models suggests the compiler's optimizations are architecture-dependent, so teams should benchmark their specific model before committing. Teams self-hosting or choosing an inference provider for Kimi, GLM, DeepSeek, or MiniMax models can use these figures to compare serving costs. Together's pitch targets the growing ecosystem serving Chinese open-weight models on NVIDIA's B-series hardware.

#Together AI#Kimi K3#DeepSeek#inference optimization
Event
Official accountTogether AI@togethercomputeSource