aifollow.news
Back IT之家 科技新闻
IT之家 科技新闻· · Original publication time AI score53

Moonshot AI releases Kimi K3.1 Pro model weightsMachine translation

月之暗面开放 Kimi K3.1 Pro 模型权重

AI-assisted summary

Moonshot AI has released the Kimi K3.1 Pro weights on Hugging Face. The model ranked first among open-weight models on Agent Arena, based on more than 17,500 real agent-task sessions. Its BF16 weights occupy about 5.6TB, making local deployment on consumer GPUs impractical.

Article · Original

The article text is unavailable in this language; an existing version is shown.

IT之家 11 月 5 日消息,按照 10 月 24 日发布时给出的承诺,月之暗面(Moonshot AI)今日在 Hugging Face 平台开放了 Kimi K3.1 Pro 的模型权重。

在 Arena 运营的 Agent Arena 智能体排行榜中,Kimi K3.1 Pro(Max)位列开放权重模型第一、全部参评模型第 13 位,并在工作、编程与聊天三个类别中均居开放模型首位。

Agent Arena 的评测方式与传统的成对盲测不同。它不要求用户对两个回答投票,而是从真实用户的智能体会话中提取信号,包括用户对任务的明确确认、表扬与投诉、纠正指令以及 Shell 命令执行错误等,再通过 Arena 称为“因果追踪”的随机化推断方法估算各模型相对于平均水平的净改进率。

Kimi K3.1 Pro 此次排名基于全球用户超过 1.75 万次真实长周期智能体任务会话,净改进率为 +5.1%,较上一代 K3(Max)的 +0.9% 提升 4.2 个百分点。

分项指标呈现出更具体的改善幅度。衡量 Shell 命令执行可靠性的 Bash Reliability 从 K3 的-4.5%(第 13 名)升至 +7.9%,在开放模型中排名第一;反映用户满意度的 Praise vs Complaint 指标从 +1.3% 升至 +7.7%;衡量模型根据用户反馈调整执行方向的 Steerability 指标达到 +6.5%,同样居开放模型首位。用户明确确认任务完成的 Confirmed Success 为 +3.9%,开放模型第四。

Kimi K3.1 Pro 采用混合专家架构,总参数量 2.8 万亿,单次推理激活约 1040 亿参数,支持约 100 万 Token 上下文窗口。

与 K3 相比,K3.1 Pro 以 K3 为基座进行了大规模联合强化学习后训练,覆盖数百个下游领域,并将部分全注意力层替换为门控稀疏注意力层,以降低长上下文建模的开销。

值得一提的是,开源权重的“可下载”与“可本地运行”之间存在显著距离。2.8 万亿参数以 BF16 精度存储的权重文件约 5.6TB,即使采用 INT4 量化也仍需约 1.4TB,单机消费级显卡基本不具备部署可能。

在终端智能体基准 Terminal-Bench 4.0 上,K3.1 Pro 得分 65.7%,高于 GPT-6 Astra 的 58.2%,与 Claude Opus 5.5 的 66.4% 相差不到 1 个百分点。

BenchmarkKimi K3.1 Pro

(max)Claude Fable 5.1

(max, w/ fallback)GPT-6 Astra

(max, w/ fallback)Claude Opus 5.5

(max, w/ fallback)DeepSeek-V4.1-Flash-1017

(max)GLM-5.3

(max)

Coding

Terminal-Bench 4.065.757.958.266.442.341.8

Terminal-Bench-Science 0.138.140.068.163.315.78.1

SWE Marathon v1.151.745.642.546.340.042.5

BenchCAD w/o tools72.860.6—73.041.2—

BenchCAD w/ tools92.792.695.996.281.6—

Program Bench31.553.550.065.029.119.0

DeepSWE v1.175.567.474.174.277.066.9

PostTrain Bench v1.236.544.641.943.830.534.5

Kimi Code Bench 2.0 (Internal)77.1———76.161.7

Kimi Code Bench 3.0 Long-horizon (Internal)61.258.866.558.551.545.5

Agentic

HLE-Full w/ tools70.065.657.267.758.862.5

BrowseComp92.1—91.590.8——

BioMysteryBench80.681.678.981.980.079.6

CharXiv (RQ) w/ python93.492.792.893.885.6—

MMMU-Pro w/ python90.088.487.889.882.5—

GDP.pdf w/ python85.783.885.583.973.0—

ChartoGraphy w/ python90.288.491.089.083.0—

BabyVision w/ python95.495.196.697.792.5—

ZeroBench_main w/ python (pass@5)63.059.058.061.051.0—

Agents' Last Exam (ALL)43.2—41.842.637.726.2

Agents' Last Exam (CLI)44.2—44.240.736.029.4

Automation Bench-AA73.359.468.569.566.662.2

GDPval-AA v2.1 (Elo-score)184617541573186617461650

AA-Briefcase v1.1 (Elo-score)167816721569180315701507

APEX-Agents69.468.664.773.569.856.6

OfficeQA Pro v264.454.554.462.240.431.1

OfficeQA Pro67.468.666.2—62.458.6

Frontier Finance58.660.845.056.754.557.5

Harvey Legal Agent23.719.1——21.0—

HANDBOOK.md44.638.8—41.528.920.8

OSWorld V2.179.680.775.881.869.3—

Business Operation (Internal)37.830.732.3—8.918.6

DECK-Arena (Internal, Elo-score)1456—1423125312311119

Design Bench (Internal, Elo-score)957—1076927863718

Kimi Hallu Bench Hard w/ tools (Internal, lower is better)8.323.35.515.023.322.6

Reasoning & Knowledge & Multimodal

GDP.pdf34.627.634.230.617.211.2

HLE-Full51.157.653.659.842.643.0*

HLE-Diamond39.455.165.860.921.625.1*

GPQA-Diamond94.293.796.193.992.291.7

CharXiv (RQ)89.088.890.092.177.0—

MMMU-Pro84.186.386.987.780.6—

ZeroBench_main (avg@5 / pass@5)29.8 / 39.023.0 / 35.043.2 / 52.035.0 / 43.011.8 / 23.0—

PerceptionBench68.363.077.672.147.7—

Kimi Hallu Bench w/o tools (Internal, lower is better)5.411.44.26.813.114.8

Cyber Security

ExploitBench74.783.010091.0—54.4

ExploitGym20.730.442.434.5—15.0

xdev-bench-hard (Internal)81.0———85.757.1

相关阅读:

《Kimi K3.1 Pro 正式发布:月之暗面最强 3 万亿级模型,编程智能体测评超 GPT-6》

《Kimi 就会员使用 K3.1 Pro 出现限流等问题致歉,为所有订阅会员进行重置》

Found an error? Send a correction