7月27日深夜，月之暗面（Moonshot AI）将Kimi K3的完整模型权重上传至Hugging Face开源仓库。2.8万亿参数，全球开源模型有史以来最大。技术报告同步发布于GitHub，采用Kimi K3 License定制许可证（类Modified MIT）。
K3是全球首个真正落地的3万亿参数级别MoE（混合专家）开源模型，激活参数约1040亿。作为对比，上一代开源最大模型DeepSeek-V3是6710亿总参数、370亿激活参数。K3在参数规模上直接跳升了一个数量级。
为什么选择现在开源
一个有趣的时间线：7月16日发布K3，7月20日官方因算力紧张暂停K3 Plus的Plus服务扩容，7月27日开放权重下载。这三件事放在一起看，逻辑就清晰了。
第一，释放算力压力。开源权重后，有能力的企业和开发者可以自己部署，减少对月之暗面API的依赖，缓解官方算力紧张的局面。
第二，争夺生态主导权。最大开源模型的名头本身就是最好的品牌建设。开发者用K3做fine-tune、做二次开发，生态就建立起来了。
第三，配合融资节奏。当开源社区都在讨论K3的时候，下一轮融资的估值故事就有了新素材。
"2.8万亿参数的模型扔上Hugging Face，就像有人把一颗小型核弹的设计图纸放到了GitHub上。能不能造出来是一回事，但图纸免费了——这件事本身就在改变游戏规则。"—— Dawn Vision编辑部
对开源生态的影响
K3开源最大的冲击对象不是OpenAI和Anthropic，而是中间层闭源模型公司。以前你可以说"我有一个比开源强的闭源模型，收API费用"，现在K3把开源模型的上限拉到了2.8万亿，你的卖点还有多少？
对云厂商来说是利好——更多企业需要部署超大模型，GPU需求会进一步增长。对中小开发者来说可能是坏消息——2.8万亿参数的模型不是一台消费级GPU能跑的，开源不等于免费可用。
但无论如何，把2.8万亿参数的模型开放下载，这件事本身就推动了整个行业的水位线向上走。
明天见。
Sources · 参考来源
声明：本文为 Dawn Vision 基于公开信息的二次创作与独立分析，仅供参考。
本文基于 Dawn Vision 认知引擎处理的 7 个源信号生成，经编辑部人工审核。素材来源：电子产品世界、Digitimes。
相关入库笔记：Kimi K3 · 月之暗面 · 开源模型 · 2.8万亿参数 · MoE · Hugging Face
Late on July 27, Moonshot AI uploaded the full model weights of Kimi K3 to Hugging Face. At 2.8 trillion parameters, it's the largest open-source model ever released. A technical paper was published simultaneously on GitHub under the Kimi K3 License (Modified MIT-style).
K3 is the world's first truly delivered 3-trillion-parameter class MoE (Mixture of Experts) open model, with ~104B activated parameters. For context, the previous largest open model, DeepSeek-V3, had 671B total / 37B activated parameters. K3 jumps an order of magnitude in parameter scale.
Why Open-Source Now
An interesting timeline: K3 announced July 16, K3 Plus service scaling paused July 20 due to compute constraints, weights open-sourced July 27. Put together, the logic becomes clear.
First, relieving compute pressure. With open weights, capable enterprises and developers can deploy on their own, reducing API load on Moonshot's infrastructure.
Second, ecosystem dominance. The title of "largest open-source model" is itself the best branding. Developers fine-tuning and building on K3 grows the ecosystem.
Third, fundraising rhythm. When the open-source community is all talking about K3, the valuation story for the next funding round writes itself.
"Dropping a 2.8-trillion-parameter model on Hugging Face is like uploading a small nuclear bomb blueprint to GitHub. Whether you can build it is one thing — but the blueprint being free changes the game."—— The Dawn Vision Editorial Desk
Impact on Open-Source Ecosystem
K3's biggest impact isn't on OpenAI or Anthropic — it's on middle-tier closed-source model companies. Previously you could sell "a better closed model than open-source" as an API. Now K3 has pushed the open-source ceiling to 2.8T — how much of a selling point is left?
Good news for cloud vendors — more enterprises need to deploy giant models, GPU demand grows. Mixed news for small developers — a 2.8T-parameter model doesn't run on a consumer GPU. Open source ≠ free to run.
But regardless, open-sourcing a 2.8T-parameter model raises the water line for the entire industry.
See you tomorrow.
Sources · 参考来源
声明：本文为 Dawn Vision 基于公开信息的二次创作与独立分析，仅供参考。
Processing 7 source signals with editorial review. Sources: EEWorld, Digitimes.
Notes: Kimi K3 · Moonshot AI · open-source · 2.8T parameters · MoE · Hugging Face
2.8万亿参数的模型扔上Hugging Face，就像有人把一颗小型核弹的设计图纸放到了GitHub上。能不能造出来是一回事，但图纸免费了——这件事本身就在改变游戏规则。
—— Dawn Vision编辑部
Dropping a 2.8-trillion-parameter model on Hugging Face is like uploading a nuclear bomb blueprint to GitHub. Whether you can build it is one thing — but the blueprint being free changes the game.
—— The Dawn Vision Editorial Desk
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 7 个源信号生成，经编辑部人工审核。素材来源：电子产品世界、Digitimes。
Processing 7 source signals with editorial review. Sources: EEWorld, Digitimes.
