📰 The InformationThe InformationCodex · 实盘不做把握 6/10
北京时间·🕒 发布 15:16:34时区推断·⟳ 抓取 15:17:12·⚡ 决策 07-17 15:19:48 +2m36s·102.2 天前
Moonshot AI 发布 Kimi K3，称开源模型编码能力接近美国前沿模型
Kimi K3 是未上市公司 Moonshot AI 的模型发布，事件方向对中国开源模型生态偏正面，但没有直接可交易主体。
📄 原文The Information↗ 原文
Moonshot AI’s New Kimi K3 Challenges U.S. Frontier Models
Chinese startup Moonshot AI’s latest large language model is sending shockwaves through the global AI industry because…
不做 阿里巴巴（09988.HK）
不做：Kimi K3 利好主体是未上市 Moonshot AI，阿里巴巴只是间接投资/生态联想，未披露订单或收入增量，且对超大市值主标的不足以形成短线正期望。
不做 美团（03690.HK）
不做：美团相关敞口仅为投资平台层面的间接参与，事件不改变美团主业收入、利润或交易量，缺乏可回测交易边界。
不做 腾讯控股（00700.HK）
不做：腾讯与 Moonshot 的历史投资传闻不足以对应本次 Kimi K3 商业化收益，且中国模型竞争加剧对其自有 AI 业务方向并不单边。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
消息传播链
- Kimi 官方文档确认 Kimi K3 为 2.8 万亿参数旗舰模型，并称完整权重将于 2026-07-27 前发布。
- Axios 报道 Kimi K3 在开源模型和成本优势上引发 AI 行业关注，说明事件已进入英文科技媒体传播。
- The Information 进一步强调 Kimi K3 在编码榜单接近美国前沿模型，引发中美模型差距收窄讨论。
关键理由
- Kimi K3 是未上市公司 Moonshot AI 的模型发布，事件方向对中国开源模型生态偏正面，但没有直接可交易主体。
- 阿里巴巴、腾讯、美团等即使存在投资或生态关联，也只是间接持股/主题敞口，未披露订单、收入分成或云推理增量，难以形成短线可回测交易。
- 该模型同时可能与阿里 Qwen、百度文心等上市公司自有模型形成竞争，上市公司影响方向并不单边。
主要风险
- 若后续出现 Moonshot 与某上市云厂、芯片厂或推理服务商的明确商业合作/采购金额，应由新信号重新评估。
- Moonshot 潜在港股 IPO 预期可能带来一级市场估值重定价，但当前不是已上市股票交易催化。
反向证据
- 做多前查反向 · 来源 web · 日期 2026-07-17 · 影响 weakens
未发现本次 Kimi K3 发布对应阿里巴巴、腾讯或美团的订单、收入分成、云服务采购或利润增量披露。
检索:Moonshot AI Kimi K3 Alibaba Tencent Meituan revenue partnership commercial agreement 原文 ↗
原文全文
展开原文 · 2,126 字
Chinese startup Moonshot AI’s latest large language model is sending shockwaves through the global AI industry because of its frontier-level coding capabilities on some benchmarks. The new model, Kimi K3, is stirring debates whether the gap between Chinese open-source and U.S. frontier models labs is fast diminishing or even being closed.
K3, unveiled on Thursday, is the world’s largest open-source model on the market, with 2.8 trillion parameters. The model now [ranks No. 1](https://x.com/arena/status/2077824029126504525?s=20) on the Frontend Code Arena coding leaderboard, above Anthropic’s most powerful model, Claude Fable 5. K3 is suitable for handling tasks that involve reasoning and knowledge work, as well as complex coding tasks that require continuous engineering for many hours, Moonshot said.
Other leaderboards show that K3 approaches the most advanced U.S. models, but doesn’t quite match them. On Artificial Analysis Intelligence Index, a comprehensive benchmark for AI models’ capabilities across maths, science, coding and reasoning, K3 [ranks No. 3](https://x.com/ArtificialAnlys/status/2077832874183860404?s=20), behind Fable 5 and OpenAI’s GPT-5.6 Sol model, but ahead of Claude Opus 4.8, which was released in May.
Earlier this year, Anthropic [accused](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks) Moonshot and other Chinese AI labs of illicitly distilling Claude models to train and improve their own models. But if K3’s performance actually rivals top Anthropic models as some benchmarks suggest, distillation may no longer explain the rapid advancements of Chinese models.
Moonshot’s launch of K3 accelerates the momentum of China’s open-source AI models, which are expanding rapidly by offering increasingly strong performance at much lower prices compared to U.S. proprietary models from Anthropic and OpenAI. Among other Chinese competitors, Beijing-based Zhipu AI’s GLM-5.2 model, released last month, has also gained popularity among developers around the world.
One coding leaderboard ranks the new Chinese open-source model above Anthropic’s Fable 5.
K3, unveiled on Thursday, is the world’s largest open-source model on the market, with 2.8 trillion parameters. The model now [ranks No. 1](https://x.com/arena/status/2077824029126504525?s=20) on the Frontend Code Arena coding leaderboard, above Anthropic’s most powerful model, Claude Fable 5. K3 is suitable for handling tasks that involve reasoning and knowledge work, as well as complex coding tasks that require continuous engineering for many hours, Moonshot said.
Other leaderboards show that K3 approaches the most advanced U.S. models, but doesn’t quite match them. On Artificial Analysis Intelligence Index, a comprehensive benchmark for AI models’ capabilities across maths, science, coding and reasoning, K3 [ranks No. 3](https://x.com/ArtificialAnlys/status/2077832874183860404?s=20), behind Fable 5 and OpenAI’s GPT-5.6 Sol model, but ahead of Claude Opus 4.8, which was released in May.
Earlier this year, Anthropic [accused](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks) Moonshot and other Chinese AI labs of illicitly distilling Claude models to train and improve their own models. But if K3’s performance actually rivals top Anthropic models as some benchmarks suggest, distillation may no longer explain the rapid advancements of Chinese models.
Moonshot’s launch of K3 accelerates the momentum of China’s open-source AI models, which are expanding rapidly by offering increasingly strong performance at much lower prices compared to U.S. proprietary models from Anthropic and OpenAI. Among other Chinese competitors, Beijing-based Zhipu AI’s GLM-5.2 model, released last month, has also gained popularity among developers around the world.
One coding leaderboard ranks the new Chinese open-source model above Anthropic’s Fable 5.
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
这条决策早于决策账本(2026-10-08 起逐条追加),没有追加记录。