开源 AI · 2026年10月

Kimi K3.1 Pro 登顶：开源 AI 与顶尖闭源模型的差距已缩短至一个月

以往，最强的开源模型通常落后最顶尖闭源模型约三个月的时间。Kimi 最新发布的模型将这一代差缩短到了 31 天，并使中国实验室在全球开源 AI 榜首的连续统治记录延长至第 765 天。

作者：ModelCompare · 2026年10月23日

10 月 23 日，Kimi 正式发布了 K3.1 Pro。该模型在 Artificial Analysis 智能指数（Artificial Analysis Intelligence Index，涵盖推理、知识、代码及 Agent 任务的高难度综合测试）中获得了 55.1 分。这一成绩与当今综合能力最强的闭源模型 Claude Opus 5.5 仅相差 2.5 分，且比此前最顶尖的开源模型——小米的 MiMo-V2.6-Pro 高出 8.8 分。Kimi 表示模型权重将于 11 月 4 日前开放免费下载；在此之前，开发者可通过 Kimi 自有 API 进行调用。

按发布时间推算，这使得开源 AI 仅落后顶尖水平约一个月：Claude Opus 5.5 是在 31 天前首次达到这一高度。而在今年 9 月小米推出 MiMo-V2.6-Pro 时，开源与闭源的差距还是 104 天（我们曾在本月早些时候对此进行过详细分析）。目前，整个 AI 领域仅有 Claude Opus 5.5（57.6 分）、Claude Sonnet 5.5（56 分）和 Muse Wonder Full（55.8 分）得分高于该模型。

在公布公开 API 价格的模型中，与 Kimi K3.1 Pro 最接近的是 Anthropic 的中型主力模型 Claude Sonnet 5.5（在最高 effort 设定下）。Kimi 新模型得分仅低 0.9 分，且标价完全相同：每百万输入 Token 2 美元，每百万输出 Token 10 美元。实际上，它相当于一款“开源版的 Sonnet 5.5”：一旦权重正式放出，任何人都可以自由运行部署。

对比 Kimi K3.1 Pro 与 Claude Sonnet 5.5

最强闭源 AI 模型与最强开源 AI 模型演进趋势

最强闭源模型 | 最强开源模型
01020304050602025Jul2026Jul

图注：基于 Artificial Analysis 智能指数（当前版本）统计各发布节点最强的闭源与开源权重模型。此前从未有开源模型夺得过总榜榜首，阴影部分表示闭源模型的领先幅度。可悬停或点击查看对应模型。

新的开源霸主来自中国，这早已不再是新闻。自 2024 年 9 月 19 日阿里巴巴的 Qwen2.5 Instruct 72B 超越 Meta 的 Llama 3.1 以来，中国实验室已经连续 765 天蝉联全球最强开源模型桂冠。而在 2023 年，“开源大模型”几乎曾完全等同于 Llama。

各实验室最强开源 AI 模型分布（按智能指数得分）

2026年10月榜首：Kimi K3.1 Pro（Kimi，55.1 分）

• Meta（美国） 38.9 Muse Spark Open
• 阿里巴巴（中国） 39.9 Qwen3.8 2.4T A95B
• Mistral（法国）
• DeepSeek（中国） 39.5 DeepSeek V4.1 Flash
• OpenChat（中国）
• 微软（美国）
• xAI（美国）
• Databricks（美国）
• Snowflake（美国）
• Nous Research（美国）
• Ai2（美国）
• Liquid AI（美国）
• 英伟达（美国）
• Perplexity（美国）
• MiniMax（中国） 29.2 MiniMax-M3
• Kimi（中国） 55.1 Kimi K3.1 Pro
• LG AI Research（韩国）
• 智谱 AI / Z.ai（中国） 44.8 GLM-5.3
• OpenAI（美国）
• ServiceNow（美国）
• InclusionAI（中国） 41.1 Ling 3.1 Flash
• Prime Intellect（美国）
• 小米（中国） 46.3 MiMo-V2.6-Pro
• MBZUAI（阿联酋） 30.5 K2 Horizon 375B A23B
• 阶跃星辰 / StepFun（中国）
• Google（美国）
• 腾讯（中国）
• Nex AGI（中国）
• Motif（韩国） 33.6 Motif 3
• Thinking Machines（美国）
• Multiverse（西班牙）

图注：仅限开源权重模型。展示自 Llama 2 发布（2023年7月18日）以来，各实验室截至相应日期发布的最高分开源模型，分数依据 Artificial Analysis 智能指数（当前版本、最高 effort 设定）。坐标轴随得分提升动态缩放，仅展示排名前 10 的机构。

六家实验室，同一个国家

开源霸主地位并非由某一家巨头独占。阿里巴巴、DeepSeek、Kimi、智谱 AI（Z.ai）、MiniMax 以及小米都曾问鼎榜首。其中，DeepSeek 领跑时间最长，累计达 307 天，这主要归功于其推理模型 R1——该模型曾在 2025 年 1 月一度引发英伟达股价震荡。打造 GLM 系列的智谱 AI（Z.ai）以 198 天的领先天数位居第二。自 2025 年 9 月以来，开源王座已发生 14 次易主，平均每 30 天就有一次更替，且没有任何一款模型能霸榜超过 72 天。

最强开源 AI 模型逐日演进路线

Meta（美国）| 阿里巴巴（中国）| DeepSeek（中国）| Kimi（中国）| 智谱 AI / Z.ai（中国）| MiniMax（中国）| 小米（中国）

• Qwen2.5
• DeepSeek R1
• DeepSeek R1 0528
• GLM-5

2025 - 2026

图注：2024 年 9 月 1 日至 2026 年 10 月 23 日期间，各时间节点 Artificial Analysis 智能指数得分最高的开放权重模型。

美国开源生态已落后三个月

10 月 15 日，Meta 发布了 Muse Spark Open，这是其自 2025 年 4 月推出 Llama 4 以来的首款开源模型，得分为 38.9。然而，中国实验室早在 91 天前发布的 Kimi K3 就已超越了这一分数。即便如此，Muse Spark Open 仍是非中国区本土研发的最强开源模型，排在韩国的 Motif 3（33.6 分）之前。

在此之前，差距甚至更加悬殊。2026 年 6 月，非中国区最强的开源模型为 Google 的 Gemma 4 26B A4B（轻量到可在笔记本上运行），其能力水平仅相当于中国团队提前 246 天（约八个月前）发布的技术水准。而 OpenAI 于 2025 年 8 月发布的 gpt-oss-120b，则从未登顶过该指数。

中国与全球其他机构的开源 AI 得分对比

中国实验室最强开源模型 | 全球其他地区最强开源模型
01020304050602025Jul2026Jul

图注：按发布日期排列的各阶段最佳开放权重模型在 Artificial Analysis 智能指数（当前版本）上的表现。蓝色阴影代表中国团队的领先优势；橙色区域代表 2024 年 9 月 Meta 的 Llama 3.1 Instruct 405B 保持领先的 18 天。

按时间维度测算，2025 年全球其他地区的顶尖开源模型落后中国的中位数为 109 天，而今年这一数字扩大到了 145.5 天。Meta 本次的更新，才终于将这一时间差距重新拉回到约三个月左右。

开发者能真正拥有的最强模型

待 Kimi K3.1 Pro 权重正式开放下载后，任何开发者和企业都可以自行运行、部署或提供推理服务，这将进一步驱动全行业推理成本的下行。过去两年来，属于你可以私有化拥有的最强模型一直诞生于中国；而如今，开源模型也从未像现在这样，如此逼近那些按 API 租赁的最顶尖闭源模型。

对比 GLM-5.3 与 Kimi K3

评估方法

智能评测分数统一采用 Artificial Analysis 智能指数（Intelligence Index，当前最新版本），各模型均基于最高 effort 参数设定，统计截至 2026 年 10 月 23 日的各模型发布时间。评测仅计入被 Artificial Analysis 标记为“开放权重”（Open Weights）的模型。本次统计快照截取自 10 月 18 日；针对 2026 年 10 月 23 日发布的 Kimi K3.1 Pro（指数得分为 55.1），鉴于 Kimi 已明确承诺在 2026 年 11 月 4 日前公开权重，故将其计入开源范畴。10 月 18 日之后发布的其他未公开权重模型不纳入本次统计。所有模型均自其发布之日起参与排名。时间轴自 2024 年 9 月开始正式统计，因在此之前各模型在当前指数下的得分普遍在 5 至 9 分之间，分差过小无法形成有效区分。整个技术竞赛起点溯源至 2023 年 7 月 18 日 Llama 2 的推出，并展示各实验室截至各时间点发布的最强开源模型。实验室按总部所在国家划分；微调模型计入发布该微调版本的机构，无论其基座模型源自何方。

---

延伸阅读

• 中美对决：两年来中国持续领跑全球最强开源 AI 模型
• 本地运行指南：你的 5 年机龄 MacBook 性能已超越免费版 ChatGPT
• 顶尖 AI 实验室竞逐：谁能在闭源赛道逼近 Claude 与 GPT 系列

Open-source AI · October 2026

Kimi K3.1 Pro puts open-source AI a month behind the best

The best open model used to trail the best closed one by about three months. Kimi’s new model cuts that to 31 days, and keeps a Chinese lab at the top of open-source AI for the 765th day in a row.

By ModelCompare · 23 October 2026

On 23 Oct, Kimi released K3.1 Pro. It scores 55.1 on the Artificial Analysis Intelligence Index, a blend of hard tests in reasoning, knowledge, coding and agent work. That is 2.5 points short of Claude Opus 5.5, the best model of any kind, and 8.8 points above the previous best open model, Xiaomi’s MiMo-V2.6-Pro. Kimi says the weights will be free to download by 4 Nov; until then it runs on Kimi’s own API.

In time, that puts open-source AI about a month behind: Claude Opus 5.5 first scored that high 31 days earlier. When Xiaomi released MiMo-V2.6-Pro in September, the gap was 104 days (we wrote about it earlier this month). Only Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56) and Muse Wonder Full (55.8) score higher today.

The closest one with a public API price is Claude Sonnet 5.5, Anthropic’s mid-sized model, at its highest effort setting. Kimi’s new model scores 0.9 points less and has the same list price: $2 per million input tokens and $10 per million output tokens. It is, in effect, an open-source Sonnet 5.5: once the weights are out, anyone can run it.

Compare Kimi K3.1 Pro with Claude Sonnet 5.5

The best closed AI model and the best open-source oneBest closed modelBest open-source model

01020304050602025Jul2026Jul

Artificial Analysis Intelligence Index (current version) of the best model released by each date, closed and open-weights. No open model has ever had the top score; the shading is the closed lead. Hover or tap for the models.

That the new leader is Chinese is not news. A Chinese lab has had the best open model every day since 19 Sep 2024, when Alibaba’s Qwen2.5 Instruct 72B passed Meta’s Llama 3.1: 765 days in a row. In 2023, open source meant Llama. Press play to watch the chart turn over.

Each lab’s best open-source AI model, by Intelligence Index score

Oct 2026Top: Kimi K3.1 Pro
Kimi, 55.1

MetaUS38.9 Muse Spark Open

AlibabaChina39.9 Qwen3.8 2.4T A95B

MistralFrance****

DeepSeek China39.5 DeepSeek V4.1 Flash

OpenChatChina****

Microsoft US****

xAI US****

Databricks US****

Snowflake US****

Nous Research US****

Ai2 US****

Liquid AI US****

NVIDIA US****

Perplexity US****

MiniMax China29.2 MiniMax-M3

KimiChina55.1 Kimi K3.1 Pro

LG AI ResearchSouth Korea****

Z.ai China44.8 GLM-5.3

OpenAIUS****

ServiceNow US****

InclusionAI China41.1 Ling 3.1 Flash

Prime IntellectUS****

Xiaomi China46.3 MiMo-V2.6-Pro

MBZUAIUAE30.5 K2 Horizon 375B A23B

StepFunChina****

Google US****

Tencent China****

Nex AGI China****

Motif South Korea33.6 Motif 3

Thinking MachinesUS****

Multiverse Spain****

Open-weights models only. The best score among open models each lab had released by that date, since Llama 2 (18 Jul 2023), on the Artificial Analysis Intelligence Index (current version, highest effort setting). The axis rescales as scores rise; the top 10 labs are shown.

Six labs, one country

The lead doesn’t belong to one company. Alibaba, DeepSeek, Kimi, Z.ai, MiniMax and Xiaomi have each held it. DeepSeek had it longest, 307 days in total, most of it with R1, the reasoning model that briefly knocked Nvidia’s share price in January 2025. Z.ai, the lab behind GLM, is second with 198. Since September 2025 there have been 14 different No. 1s, about one every 30 days, and none lasted more than 72 days.

Who had the best open-source AI model, day by dayMeta (US)Alibaba (China)DeepSeek (China)Kimi (China)Z.ai (China)MiniMax (China)Xiaomi (China)

• • Qwen2.5
• • DeepSeek R1
• DeepSeek R1 0528
• • • • • • • • • GLM-5

---

20252026

Hover or tap a stretch to see the model.

The open-weights model with the highest Artificial Analysis Intelligence Index score among those released by each date, 1 Sep 2024 to 23 Oct 2026.

America is three months behind

On 15 Oct, Meta released Muse Spark Open, its first open model since Llama 4 in April 2025. It scores 38.9. A Chinese lab first beat that 91 days earlier, with Kimi K3. That makes it the best open model from anywhere outside China, ahead of South Korea’s Motif 3 (33.6).

It had been worse. In June 2026, the best non-Chinese open model was Google’s Gemma 4 26B A4B, a model small enough for a laptop, and it matched what China had released 246 days, or about eight months, earlier. OpenAI’s gpt-oss-120b, in August 2025, never topped this index.

The best open-source AI score from Chinese labs, and from everyone elseBest open model from ChinaBest open model from anywhere else

01020304050602025Jul2026Jul

Artificial Analysis Intelligence Index (current version) of the best open-weights model released by each date. Blue shading is China’s lead; orange, the 18 days in September 2024 when Meta’s Llama 3.1 Instruct 405B led. Hover or tap for the models.

Measured in time, the rest of the world’s best open model was a median of 109 days behind China’s in 2025 and 145.5 days this year. Meta’s release brings it back to about three months.

Liked this one?

We run experiments on AI models that make you think. Get the next one by email. Unsubscribe anytime.

The best model you can own

Once Kimi K3.1 Pro’s weights are out, anyone can run it or serve it, which keeps prices down. For two years the best model you could own has come from China. It has never been this close to the best model you can rent.

Compare GLM-5.3 with Kimi K3

Method

Intelligence scores are the Artificial Analysis Intelligence Index (current version), taking each model at its highest effort setting, by release date, as of 23 Oct 2026. Only models Artificial Analysis marks as open weights count. Our snapshot is from 18 October; we added Kimi K3.1 Pro, released on 23 Oct 2026 with an index score of 55.1, and count it as open because Kimi says it will publish the weights by 4 Nov 2026. Other models released after 18 October aren’t counted. A model counts from its release date. We count from September 2024 because before that every model scores between 5 and 9 on the current index, too close to rank. The race starts at Llama 2 on 18 Jul 2023 and shows each lab’s best open model on each date. Labs are placed by where they are based; a fine-tune counts for the lab that released it, wherever its base model came from.

ModelCompare sends one prompt to up to 6 models at once, open and closed side by side, and shows each answer with what it cost. Sign up with Google and you get $0.10 of free credits to start, with no API keys:

compare models now →

China vs the USChina has made the best open-source AI model every day for two years765 days in a row, 15 handoffs between Chinese labs, and how far behind America's best open model is.Read the postRun it yourselfYour 5-year-old MacBook can now beat free ChatGPTThe best open model a 32GB laptop can run, against the free ChatGPT, since 2024.Read the postAI labsEveryone else is three months behind Claude and GPTThe same race with closed models in it: who has had the best AI model on each day since GPT-4.Watch the race

Liked this one?

We run experiments on AI models that make you think. Get the next one by email. Unsubscribe anytime.