Is Kimi K3 open-sourced, and what are its parameters and license? Kimi K3 开源了吗?参数量和许可是什么?

Yes. On July 27, 2026, Moonshot AI released Kimi K3's full weights under a modified MIT license: 2.8T total / ~104B active parameters, ~1.56TB of files, plus three training-infra components (MoonEP, FlashKDA, AgentEnv). It topped Hugging Face's trending chart within 30 minutes with 4,000+ likes, with Day-0 support from Huawei Ascend and Alibaba Cloud and integrations from Cursor and Cognition. 已开源。月之暗面于 2026 年 7 月 27 日以修改版 MIT 许可开源 Kimi K3 全部权重:总参数 2.8 万亿、激活约 1040 亿,权重文件约 1.56TB,并同步开源 MoonEP、FlashKDA、AgentEnv 三项训练 Infra。上线 30 分钟即以超 4000 赞登顶 Hugging Face 趋势榜,华为昇腾、阿里云等完成 Day0 适配,Cursor、Cognition 宣布接入。

When will DeepSeek IPO, and what is its latest valuation? DeepSeek 什么时候 IPO?最新估值多少?

Per Bloomberg (July 14) and WSJ (July 15) reports, DeepSeek has begun onshore IPO preparations for a Shanghai listing, possibly filing this year with a 2027 debut. Valuation rose from ~$52B post-money in June's first round to ~$71B pre-money (~RMB 480B) in July talks; on July 26 it reportedly paused new fundraising to prioritize the IPO. The SSE issued a guideline on June 17 applying the STAR Market's fifth listing standard to large-model companies. 据彭博(7 月 14 日)与 WSJ(7 月 15 日)报道,DeepSeek 已启动境内 IPO 准备,拟在上海上市,或年内递表、2027 年挂牌。估值从 6 月首轮融资投后约 520 亿美元升至 7 月谈判投前约 710 亿美元(约 4800 亿元);7 月 26 日据报道暂缓新融资、优先推进 IPO。上交所 6 月 17 日已发布 AI 大模型企业适用科创板第五套标准的审核指引。

How did Hong Kong-listed Zhipu and MiniMax perform after lock-up expiry? 港股 AI 股智谱、MiniMax 解禁后表现如何?

Highly volatile. Zhipu rose 13.35% on its first lock-up expiry day (July 8), reportedly adding over HK$100B in market value; MiniMax saw ~50% of shares unlock on July 9, with founders voluntarily extending lock-ups by 12 months. MiniMax then surged 17%+ on July 27, before Zhipu fell as much as 20% intraday and MiniMax over 12% on July 28. Per EY, H1 Hong Kong IPOs averaged a 61% first-day return—re-rating via listing comes with high volatility. 波动剧烈。智谱 7 月 8 日首批解禁日反涨 13.35%,据报道单日市值增逾千亿港元;MiniMax 7 月 9 日约 50% 股份解禁,创始团队自愿延长锁定 12 个月。随后 7 月 27 日 MiniMax 涨超 17%,28 日智谱盘中一度跌 20%、MiniMax 跌超 12%。据 EY,上半年港股 IPO 平均首日回报 61%,「上市重估」伴随高波动。

← 返回研究报告

中国 AI 出海

中国 AI 出海 2026:开源突围与全球竞速(7 月中旬更新)

天际研究 · AI 研究院 · 2026.07.06 · 全文 24 页 · 预览 6 页

🎧

听报告 · 音频摘要

约 5-8 分钟 · AI 朗读 · 含摘要与全部关键发现(切换网站语言即切换中/英音频)

摘要

(数据更新至 2026-08-01)7 月下旬,中国 AI 出海从「发布竞速」进入「兑现与定价」阶段:月之暗面 7 月 27 日如约开源 Kimi K3 全部权重(2.8 万亿参数、约 1.56TB)并同步开源三项训练 Infra,上线 30 分钟登顶 Hugging Face 趋势榜;阿里在 WAIC 预览 2.4 万亿参数 Qwen3.8-Max。资本端,DeepSeek 据报道启动 IPO 准备、估值谈判升至约 710 亿美元;港股「AI 双雄」智谱与 MiniMax 经历解禁与暴涨暴跌。美国方面,苹果诉 OpenAI 案进入程序,OpenAI 官宣 1220 亿美元融资,Anthropic 最快 10 月上市。报告沿用中性价值链框架,更新开源出海、Agent 商业化、估值叙事与政策变量。

关键发现

01开权重兑现:7 月 27 日深夜月之暗面开源 Kimi K3 全部权重(修改版 MIT 许可、总参数 2.8 万亿/激活约 1040 亿、权重文件约 1.56TB),并同步开源 MoonEP、FlashKDA、AgentEnv 三项训练 Infra;上线 30 分钟以超 4000 赞登顶 Hugging Face 趋势榜(平台史上最快),华为昇腾、阿里云与 Nebius、Baseten、Fireworks 完成 Day0 适配,Cursor、Cognition(Devin)宣布接入。
02开源竞速升级:7 月 19 日阿里在上海 WAIC 预览 Qwen3.8-Max(2.4 万亿参数),自称「仅次于 Claude Fable 5」但未公布基准与模型卡,承诺「尽快」开权重(截至 8 月 1 日未兑现);7 月 24 日黄仁勋转发 a16z 发起、英伟达/Perplexity/Ollama 等联署公开信,主张前沿开源与闭源并存——开源已从中国策略升级为美国政策议题。
03资本时间表:Anthropic 6 月 1 日秘密递交 S-1 后,据彭博 7 月 15 日报道投行已安排投资者会议、最快 10 月登陆纳斯达克(5 月融资 650 亿美元、投后估值 9650 亿美元);OpenAI 7 月 22 日官宣完成 1220 亿美元融资(亚马逊、英伟达、软银锚定),其 IPO 据多家外媒报道推迟至 2027。
04DeepSeek 定价锚:彭博(7/14)与 WSJ(7/15)报道其启动境内 IPO 准备(拟上海上市、或年内递表、2027 挂牌);估值从 6 月首轮融资投后约 520 亿美元升至 7 月谈判投前约 710 亿美元(约 4800 亿元),7 月 26 日据报道暂缓新融资、优先推进 IPO;上交所 6 月 17 日已发布 AI 大模型企业适用科创板第五套上市标准的审核指引。
05港股压力测试:智谱 7 月 8 日首批解禁日反涨 13.35%、据报道单日市值增逾千亿港元;MiniMax(00100.HK,1 月 9 日上市)7 月 9 日约 50% 股份解禁、创始团队自愿延长锁定 12 个月;7 月 27 日 MiniMax 涨超 17%,28 日智谱盘中一度跌 20%、MiniMax 跌超 12%——「上市重估」路径伴随高波动(据 EY,上半年港股 IPO 平均首日回报 61%)。
06Agent 付费验证:软银 7 月 14 日成为 Sierra 日本独家渠道(此前 LINEMO 客服部署问询解决率 97%、满意度 93%),将扩展至软银与 Y!mobile 主品牌;月之暗面据报道 6 月 ARR 突破 3 亿美元,拟 8 月启动目标投前 500 亿美元的上市前融资谈判并已向投资人发送上市议案。

Executive Summary · English

(Data as of 2026-08-01) In late July, China's AI going-global story shifted from launch-racing to delivery and pricing. Moonshot AI open-sourced Kimi K3's full weights (2.8T parameters, ~1.56TB) on July 27 as promised, together with three training-infrastructure components, topping Hugging Face's trending chart within 30 minutes; Alibaba previewed the 2.4T-parameter Qwen3.8-Max at WAIC Shanghai. On the capital side, DeepSeek reportedly began IPO preparations with valuation talks near $71B, while Hong Kong-listed Zhipu and MiniMax went through first lock-up expiries and sharp price swings. In the U.S., Apple's lawsuit against OpenAI moved into procedure, OpenAI announced a $122B financing, and Anthropic could list as soon as October. The report keeps its neutral value-chain framework, updating open-source expansion, agent commercialization, valuation narratives and policy variables.

01Delivery on open weights: Late on July 27, Moonshot AI released Kimi K3's full weights (modified MIT license; 2.8T total / ~104B active parameters; ~1.56TB files) plus three training-infra components (MoonEP, FlashKDA, AgentEnv). It topped Hugging Face's trending chart within 30 minutes with 4,000+ likes—the platform's fastest launch ever—with Day-0 support from Huawei Ascend, Alibaba Cloud, Nebius, Baseten and Fireworks, and integrations announced by Cursor and Cognition (Devin).
02Open-source race escalates: On July 19 at WAIC Shanghai, Alibaba previewed Qwen3.8-Max (2.4T parameters), claiming it is 'second only to Claude Fable 5' but publishing no benchmarks or model card, with open weights promised 'soon' (not delivered as of Aug 1). On July 24, Jensen Huang amplified an a16z-initiated open letter co-signed by Nvidia, Perplexity and Ollama arguing frontier open and closed models should coexist—open source has become a U.S. policy debate, not just a Chinese strategy.
03Capital-markets timeline: After Anthropic's confidential S-1 filing on June 1, Bloomberg reported on July 15 that bankers began investor meetings for a Nasdaq listing as soon as October (post-money $965B after May's $65B Series H). OpenAI announced a $122B financing on July 22 (anchored by Amazon, Nvidia and SoftBank); its own IPO has reportedly slipped to 2027.
04DeepSeek as the pricing anchor: Bloomberg (7/14) and WSJ (7/15) reported it has begun onshore IPO preparations (Shanghai listing, filing possibly this year, 2027 debut). Valuation moved from ~$52B post-money in June's first round to ~$71B pre-money (~RMB 480B) in July talks; on July 26 it reportedly paused new fundraising to prioritize the IPO. The SSE issued guideline No.10 on June 17 applying the STAR Market's fifth listing standard to large-model companies.
05Hong Kong stress test: Zhipu rose 13.35% on its first lock-up expiry day (July 8), reportedly adding over HK$100B in market value; MiniMax (00100.HK, listed Jan 9) saw ~50% of shares unlock on July 9 with founders voluntarily extending lock-ups by 12 months. MiniMax surged 17%+ on July 27, then Zhipu fell as much as 20% intraday and MiniMax over 12% on July 28—the 're-rating via listing' path now comes with high volatility (EY: H1 average first-day HK IPO return of 61%).
06Agent monetization validated: SoftBank became Sierra's exclusive Japan channel on July 14 (after a LINEMO deployment reached 97% inquiry resolution and 93% CSAT), expanding to the SoftBank and Y!mobile brands. Moonshot reportedly crossed $300M ARR in June and plans to start pre-IPO funding talks in August targeting a $50B pre-money valuation, having reportedly circulated a listing proposal to investors.

一、开源模型:出海主通道再提速

7 月 17 日凌晨发布的 Kimi K3(2.8 万亿参数 MoE、百万 token 上下文)于 7 月 27 日深夜兑现开权重承诺:全部权重(修改版 MIT、约 1.56TB)登陆 Hugging Face,并同步开源 MoonEP、FlashKDA、AgentEnv 三项训练 Infra;上线 30 分钟以超 4000 赞登顶趋势榜,创平台最快纪录。生态呈「Day0 适配」:华为昇腾、阿里云超节点与 Nebius、Baseten、Fireworks 首日上线,Cursor、Devin 宣布接入。榜单上,K3 以 1679 Elo 登顶前端编程盲测(高于 Claude Fable 5 的 1631),但综合能力仍逊于两大闭源旗舰,API 定价约 12 美元/百万 token 亦告别「中国折扣」。竞速仍在加密:7 月 19 日阿里在 WAIC 预览 2.4 万亿参数 Qwen3.8-Max,自称仅次于 Fable 5 但未附基准,开权重承诺截至 8 月 1 日未兑现;7 月 24 日黄仁勋转发 a16z 发起的公开信,主张前沿开源与闭源并存——开源已从中国策略升级为美国政策议题。下载量大头仍是小尺寸模型,「旗舰立标杆+小模型渗透」的双层结构未变。

二、美国竞争格局:诉讼、硬件与资本军备

法务层(已证实):苹果 7 月 10 日在加州北区联邦法院起诉 OpenAI、其硬件子公司 io Products 及两名前苹果员工(现 OpenAI 硬件负责人 Tang Yew Tan 与前工程师 Chang Liu),寻求赔偿与禁令;OpenAI 否认。据报道(FT),苹果已向数十名跳槽前员工发出律师函;起诉书指控均为单方陈述,截至 7 月底法院未作实质性裁决,属存疑事项。硬件层:OpenAI 7 月 15 日发布 230 美元 Codex Micro(与 Work Louder 合作的限量编程 Agent 控制键盘),约 12 小时售罄、7 月 24 日发货,eBay 转售价一度达发行价数倍。资本层:OpenAI 7 月 22 日官宣完成 1220 亿美元融资(亚马逊、英伟达、软银锚定),并披露企业收入占比超 40%、Codex 周活跃用户超 200 万(三个月增 5 倍);其 IPO 据多家外媒报道推迟至 2027,而 Anthropic 已于 6 月 1 日秘密递交 S-1,据彭博报道投行已安排投资者会议、最快 10 月上市。模型层:马斯克 7 月 18 日称 xAI 正训练 2 万亿参数 Grok 4.6、目标直指超越 K3。竞争已同时在法庭、货架与资本表上展开。

三、一级市场:估值叙事的「溢价」结构

DeepSeek 成为 7 月中下旬的定价锚:据彭博(7 月 14 日)与 WSJ(7 月 15 日),其已启动境内 IPO 准备(拟上海上市、或年内递表、2027 挂牌),并与会计师、投行接洽;估值从 6 月首轮融资(约 70 亿美元)投后约 520 亿美元,升至 7 月新一轮谈判投前约 710 亿美元(约 4800 亿元);7 月 26 日据报道其暂缓新融资、优先推进 IPO。月之暗面从性能叙事切换到资本叙事:据每经报道,拟 8 月启动目标投前 500 亿美元的上市前融资谈判(现有轮次投前约 315 亿美元),并已向投资人发送上市议案、最快 6 个月内登陆港交所;ARR 据报道 6 月突破 3 亿美元,对应 P/ARR 仍达百倍量级。溢价结构延续原判断:DeepSeek 含技术影响力与战略稀缺性溢价,月之暗面偏财务与性能逻辑;市场为「叙事」定价的程度加深,而上交所 6 月发布的科创板第五套标准指引为这类叙事提供了境内退出通道——回调风险与制度红利同时累积。

关键问答

Kimi K3 开源了吗?参数量和许可是什么?

已开源。月之暗面于 2026 年 7 月 27 日以修改版 MIT 许可开源 Kimi K3 全部权重:总参数 2.8 万亿、激活约 1040 亿,权重文件约 1.56TB,并同步开源 MoonEP、FlashKDA、AgentEnv 三项训练 Infra。上线 30 分钟即以超 4000 赞登顶 Hugging Face 趋势榜,华为昇腾、阿里云等完成 Day0 适配,Cursor、Cognition 宣布接入。

DeepSeek 什么时候 IPO?最新估值多少?

据彭博(7 月 14 日)与 WSJ(7 月 15 日)报道,DeepSeek 已启动境内 IPO 准备,拟在上海上市,或年内递表、2027 年挂牌。估值从 6 月首轮融资投后约 520 亿美元升至 7 月谈判投前约 710 亿美元(约 4800 亿元);7 月 26 日据报道暂缓新融资、优先推进 IPO。上交所 6 月 17 日已发布 AI 大模型企业适用科创板第五套标准的审核指引。

港股 AI 股智谱、MiniMax 解禁后表现如何?

波动剧烈。智谱 7 月 8 日首批解禁日反涨 13.35%,据报道单日市值增逾千亿港元;MiniMax 7 月 9 日约 50% 股份解禁,创始团队自愿延长锁定 12 个月。随后 7 月 27 日 MiniMax 涨超 17%,28 日智谱盘中一度跌 20%、MiniMax 跌超 12%。据 EY,上半年港股 IPO 平均首日回报 61%,「上市重估」伴随高波动。

延伸观看 · 收听

🎙️AI 与关税大潮下的北美大厂见闻🎧 小宇宙 · 国内可听