广告
加载中

月之暗面发布Kimi K3.1 Pro 11月开放开源权重

亿邦AI 2026-10-26 16:13
亿邦AI 2026/10/26 16:13

邦小白快读

EN
全文速览

总1:发布与获取信息。

1.月之暗面2026年10月24日推出Kimi K3.1 Pro,为K3后训练版,参数3万亿,计划11月4日开源全部权重。

2.可通过官方网页、移动端和API访问;企业用户需管理员开启权限,API可选零数据留存。

总2:性能与价格。

1.第三方排名显示其超过小米MiMo-V2.6-Pro、DeepSeek V4.1-Flash-1017、Meta Muse Wonder和OpenAI GPT-6 Astra,待权重开放后为全球开源权重第一、全品类第四。

2.标准API每百万输入2美元、输出10美元、缓存输入0.2美元,较K3降三分之一,为Claude Opus 5.5和Meta Muse Wonder一半,仍高于DeepSeek和小米;高速版输入4美元、输出20美元,速度快2至3倍。

总3:实操与风险。

1.适合代码仓库维护、长链路编码、金融文档处理、3D原型制作、漏洞排查修复等定向测试,不建议全量替代现有模型。

2.幻觉率从53%降至44%,准确率从48%升至51%,输出需人工核验,不可直接用于生产自动决策;选型要以自身工作负载测试为准。

总1:品牌发布与定位。

1.月之暗面以K3后训练版推出K3.1 Pro,3万亿参数,11月4日开放全部权重,走开源权重与API并行路线。

2.第三方排名称其待开放后为全球开源权重第一、全品类第四,并以低于Claude Opus 5.5约65%的单任务成本强化性价比定位。

总2:定价与竞争。

1.标准API输入2美元、输出10美元、缓存输入0.2美元,较K3降三分之一,为Claude Opus 5.5和Meta Muse Wonder一半,但仍高于DeepSeek和小米;DeepSeek非工作日及非高峰还有折扣。

2.高速版输入4美元、输出20美元,速度为标准版2至3倍,形成标准与高速两档价格产品。

总3:产品研发与营销风险。

1.能力覆盖复杂编码、金融分析、3D创作、软件安全,游戏开发支持Unreal、Unity、Godot,3D可结合草图、参考图和Blender、CAD脚本。

2.社交平台出现疑似推广指引截图,要求创作者不标注付费合作,争议尚未独立验证,月之暗面未回应;按X规则和FTC指引,付费等实质性关联须披露。

总1:政策与市场机会。

1.11月4日将开放全部模型权重,暂不支持自托管,本地部署企业需等权重开放;当前API在新加坡服务器,数据合规、政府采购用户需先核验条款。

2.美国国土安全部曾将DeepSeek排除民用联邦采购并要求移除,禁令暂未涉及Kimi,但敏感领域仍需确认合规。

总2:价格与销售机会。

1.标准API每百万输入2美元、输出10美元、缓存输入0.2美元,较K3降三分之一;高速版输入4美元、输出20美元,速度为标准版2至3倍。

2.单智能指数任务平均成本约2.11美元,较Claude Opus 5.5低65%;但高吞吐常规任务DeepSeek、小米仍有价格优势。

总3:风险与可学习点。

1.适合代码仓库维护、长链路编码、金融文档处理、3D原型制作、漏洞排查修复做定向测试,不建议全量替代。

2.幻觉率从53%降至44%、准确率从48%升至51%,需人工核验;疑似推广截图要求不标付费合作,来源未验证,提示合作推广要合规披露。

总1:设计与生产机会。

1.3D创作可根据文字需求、手绘草图或参考图生成模型,通过Blender、CAD脚本完成机械、建筑、家具等品类设计调整。

2.配套3D Make应用支持在线编辑模型、导出3D打印文件及动态展示内容,利于原型制作和设计协同。

总2:数字化与电商启示。

1.游戏开发支持Unreal、Unity、Godot等引擎,可结合网页检索、3D生成产出完整交互内容,适合数字化内容生产。

2.复杂编码、多步骤工具调用和重复调试迭代提升明显,Terminal-Bench 4.0从13%升至67%,可用于代码仓库维护、长链路编码和漏洞排查修复。

总3:落地注意。

1.当前暂不支持自托管,有本地部署需求需等11月权重开放;API在新加坡服务器,数据合规和政府采购要求需核验。

2.模型幻觉率仍较高,输出需人工核验,不建议直接用于生产自动决策;应先按自身工作负载做定向测试,综合交付质量、成本和人工修正量。

总1:技术与趋势。

1.性能提升全部来自K3大规模后训练,覆盖数百任务域强化学习,优化训练软件、长上下文、算力效率和工具隔离环境,未采用全新架构。

2.模型提供low、high、max三档推理力度,默认max;切换力度会导致前缀缓存失效,温度固定1.0、top-p固定0.95,支持自动、关闭、强制工具调用模式。

总2:客户痛点与方案。

1.复杂编码、金融分析、3D创作、软件安全表现突出,Terminal-Bench 4.0达67%,CyberGym-E2E-AA达95%,可做代码仓库维护、金融文档处理、3D原型和漏洞修复。

2.多轮对话需完整保留模型返回全部信息包括推理内容才可复现测试效果;API可选零数据留存,企业需管理员开启权限。

总3:交付风险。

1.暂不支持自托管,API在新加坡服务器,数据合规和政府采购客户需核验条款。

2.幻觉率从53%降至44%,准确率从48%升至51%,输出需人工核验,不可直接用于生产自动决策;应按客户工作负载测试交付质量、成本和修正量。

总1:平台访问与运营。

1.开发者与普通用户可通过官方网页、移动端应用及API访问;企业用户需管理员手动开启模型访问权限,API用户可选零数据留存模式。

2.接口提供low、high、max三档推理力度,默认max,切换会导致前缀缓存失效;支持自动、关闭、强制工具调用模式,多轮需完整保留模型返回全部信息包括推理内容。

总2:商业需求与合规风向。

1.当前API部署在新加坡服务器,有数据合规、政府采购要求的用户需提前核验服务条款;模型暂不支持自托管,本地部署需等11月权重开放。

2.美国国土安全部曾禁令将DeepSeek排除民用联邦采购并要求移除,暂未涉及Kimi,但敏感领域平台需提前确认合规。

总3:推广与平台规则。

1.社交平台X出现疑似Kimi推广指引截图,要求创作者不得标注付费合作,并包装为对标Claude Opus 5.5的低价开源模型,来源尚未独立验证,月之暗面未回应。

2.按X平台规则及美国联邦贸易委员会指引,品牌方与推广者存在付费、免费产品服务等实质性关联时必须明确披露,平台运营需防范未披露商业关系。

总1:产业新动向与基准差异。

1.模型为3万亿参数K3后训练版,11月4日开放全部权重,第三方称待开放后为全球开源权重第一、全品类第四。

2.Terminal-Bench 4.0从13%升至67%,CyberGym-E2E-AA达95%,GDPval-AA Elo为1846、全球第三;科学计算仍低于GPT-6 Astra和Claude Opus 5.5。

总2:技术路线与商业模式。

1.提升全部来自大规模后训练和数百任务域强化学习,优化训练软件、长上下文、算力效率、工具隔离环境,未采用全新架构。

2.开源权重与API并行,标准API输入2美元、输出10美元、缓存输入0.2美元,高速版输入4美元、输出20美元;单任务平均成本约2.11美元,较Claude Opus 5.5低65%。

总3:新问题与政策启示。

1.幻觉率从53%降至44%,准确率从48%升至51%,仍需人工核验;暂不支持自托管,API在新加坡,DHS对DeepSeek的联邦采购禁令暂未涉及Kimi。

2.X平台出现疑似未披露付费合作的推广指引截图,来源未验证,按X规则和FTC指引实质性关联须披露。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Summary 1: Release and access to information.

1. Moonshot AI launched Kimi K3.1 Pro on October 24, 2026, a post-trained version of K3 with 3 trillion parameters, and plans to open-source all weights on November 4.

2. It is accessible via the official website, mobile app, and API; enterprise users need an administrator to enable permissions, and the API offers an optional zero-data-retention mode.

Summary 2: Performance and pricing.

1. Third-party rankings show it surpassing Xiaomi MiMo-V2.6-Pro, DeepSeek V4.1-Flash-1017, Meta Muse Wonder, and OpenAI GPT-6 Astra; once the weights are released, it is expected to rank first among global open-weight models and fourth overall.

2. Standard API pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens, one-third lower than K3, half the price of Claude Opus 5.5 and Meta Muse Wonder, but still higher than DeepSeek and Xiaomi. The high-speed version is $4 per million input tokens and $20 per million output tokens and is 2–3 times faster.

Summary 3: Practical use and risks.

1. It is suited for targeted testing in code repository maintenance, long-chain coding, financial document processing, 3D prototyping, and vulnerability detection and remediation; it is not recommended as a full replacement for existing models.

2. The hallucination rate fell from 53% to 44%, while accuracy rose from 48% to 51%. Outputs require human verification and cannot be used directly for automated production decisions; model selection should be based on testing against one's own workload.

Summary 1: Brand launch and positioning.

1. Moonshot AI is launching K3.1 Pro as a post-trained version of K3, with 3 trillion parameters and all weights to be opened on November 4, pursuing a parallel open-weight and API strategy.

2. Third-party rankings say that once released it will be first in global open weights and fourth overall, reinforcing its value-for-money positioning with a per-task cost about 65% lower than Claude Opus 5.5.

Summary 2: Pricing and competition.

1. Standard API pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens, one-third lower than K3 and half the price of Claude Opus 5.5 and Meta Muse Wonder, but still higher than DeepSeek and Xiaomi; DeepSeek also offers discounts on non-workdays and off-peak hours.

2. The high-speed version is $4 per million input tokens and $20 per million output tokens and is 2–3 times faster than the standard version, creating standard and high-speed pricing tiers.

Summary 3: Product R&D and marketing risks.

1. Capabilities cover complex coding, financial analysis, 3D creation, and software security; game development supports Unreal, Unity, and Godot, and 3D work can incorporate sketches, reference images, and Blender and CAD scripts.

2. Screenshots that appear to be promotional guidelines emerged on social platforms, requiring creators not to label paid partnerships; the controversy has not been independently verified, and Moonshot AI has not responded. Under X rules and FTC guidance, paid and other material connections must be disclosed.

Summary 1: Policy and market opportunities.

1. All model weights will be opened on November 4, but self-hosting is not yet supported; enterprises requiring local deployment must wait for the weight release. The current API is on Singapore servers, and data-compliance and government procurement users need to verify terms first.

2. The U.S. Department of Homeland Security previously excluded DeepSeek from civilian federal procurement and required its removal; the ban does not yet cover Kimi, but sensitive sectors still need to confirm compliance.

Summary 2: Pricing and sales opportunities.

1. Standard API pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens, one-third lower than K3; the high-speed version is $4 per million input tokens and $20 per million output tokens and is 2–3 times faster than the standard version.

2. The average cost per Intelligence Index task is about $2.11, 65% lower than Claude Opus 5.5; however, DeepSeek and Xiaomi still have price advantages for high-throughput routine tasks.

Summary 3: Risks and lessons.

1. It is suited for targeted testing in code repository maintenance, long-chain coding, financial document processing, 3D prototyping, and vulnerability detection and remediation; it is not recommended as a full replacement.

2. The hallucination rate fell from 53% to 44%, and accuracy rose from 48% to 51%, but outputs require human verification; screenshots appearing to show promotional guidance require not labeling paid partnerships, and the source is unverified, highlighting that collaborative promotions must disclose compliance.

Summary 1: Design and production opportunities.

1. 3D creation can generate models from text requirements, hand-drawn sketches, or reference images, and adjust designs for mechanical, architectural, furniture, and other categories via Blender and CAD scripts.

2. The companion 3D Make app supports online model editing, exporting 3D printing files, and dynamic display content, facilitating prototyping and design collaboration.

Summary 2: Digitalization and e-commerce implications.

1. Game development supports Unreal, Unity, Godot, and other engines, and can combine web retrieval and 3D generation to produce complete interactive content, suitable for digital content production.

2. Complex coding, multi-step tool calls, and repeated debugging iterations improved significantly; Terminal-Bench 4.0 rose from 13% to 67%, enabling use in code repository maintenance, long-chain coding, and vulnerability detection and remediation.

Summary 3: Implementation cautions.

1. Self-hosting is not yet supported; those needing local deployment must wait for the November weight release. The API is on Singapore servers, and data compliance and government procurement requirements need verification.

2. The model's hallucination rate remains relatively high, and outputs require human verification; it is not recommended for direct use in automated production decisions. Targeted testing should first be conducted against one's own workload, considering overall delivery quality, cost, and manual correction effort.

Summary 1: Technology and trends.

1. All performance gains come from large-scale post-training of K3, covering reinforcement learning across hundreds of task domains and optimizing training software, long context, compute efficiency, and tool isolation environments, without adopting a new architecture.

2. The model offers low, high, and max reasoning effort levels, with max as the default; switching effort levels invalidates the prefix cache. Temperature is fixed at 1.0 and top-p at 0.95, and it supports automatic, off, and forced tool-calling modes.

Summary 2: Customer pain points and solutions.

1. It performs strongly in complex coding, financial analysis, 3D creation, and software security; Terminal-Bench 4.0 reaches 67% and CyberGym-E2E-AA reaches 95%, enabling code repository maintenance, financial document processing, 3D prototyping, and vulnerability remediation.

2. For multi-turn conversations, all model-returned information, including reasoning content, must be fully retained to reproduce test results; the API offers an optional zero-data-retention mode, and enterprises need an administrator to enable permissions.

Summary 3: Delivery risks.

1. Self-hosting is not yet supported, and the API is on Singapore servers; data-compliance and government procurement customers need to verify terms.

2. The hallucination rate fell from 53% to 44%, and accuracy rose from 48% to 51%, but outputs require human verification and cannot be used directly for automated production decisions; delivery quality, cost, and correction effort should be tested against the customer's workload.

Summary 1: Platform access and operations.

1. Developers and general users can access via the official website, mobile app, and API; enterprise users need an administrator to manually enable model access permissions, and API users can opt for zero data retention.

2. The interface offers low, high, and max reasoning effort levels, with max as the default; switching invalidates the prefix cache. It supports automatic, off, and forced tool-calling modes, and multi-turn use requires fully retaining all model-returned information, including reasoning content.

Summary 2: Commercial requirements and compliance trends.

1. The current API is deployed on Singapore servers, and users with data compliance or government procurement requirements must verify the terms of service in advance; the model does not yet support self-hosting, and local deployment must wait for the November weight release.

2. The U.S. Department of Homeland Security previously banned DeepSeek from civilian federal procurement and required its removal; this does not yet involve Kimi, but platforms in sensitive sectors need to confirm compliance in advance.

Summary 3: Promotion and platform rules.

1. Screenshots appearing to be Kimi promotional guidelines appeared on social platform X, asking creators not to label paid partnerships and packaging it as a low-cost open-source model benchmarked against Claude Opus 5.5; the source has not been independently verified, and Moonshot AI has not responded.

2. Under X platform rules and U.S. Federal Trade Commission guidance, brands and promoters must clearly disclose paid, free product/service, and other material connections; platform operators need to guard against undisclosed commercial relationships.

Summary 1: Industry developments and benchmark differences.

1. The model is a 3-trillion-parameter post-trained version of K3, with all weights to be opened on November 4; third parties say that once released it will be first in global open weights and fourth overall.

2. Terminal-Bench 4.0 rose from 13% to 67%, CyberGym-E2E-AA reached 95%, and GDPval-AA Elo is 1846, third globally; scientific computing still lags GPT-6 Astra and Claude Opus 5.5.

Summary 2: Technology roadmap and business model.

1. All improvements come from large-scale post-training and reinforcement learning across hundreds of task domains, optimizing training software, long context, compute efficiency, and tool isolation environments, without adopting a new architecture.

2. Open weights and API are pursued in parallel; standard API pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached input tokens, while the high-speed version is $4 per million input tokens and $20 per million output tokens. The average cost per task is about $2.11, 65% lower than Claude Opus 5.5.

Summary 3: New issues and policy implications.

1. The hallucination rate fell from 53% to 44% and accuracy rose from 48% to 51%, but human verification is still required; self-hosting is not yet supported, the API is in Singapore, and the DHS federal procurement ban on DeepSeek does not yet cover Kimi.

2. Screenshots appearing to show promotional guidelines for undisclosed paid partnerships appeared on X; the source is unverified, and under X rules and FTC guidance, material connections must be disclosed.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年10月24日,中国AI创业公司月之暗面推出最新大语言模型Kimi K3.1 Pro。该模型为今年7月发布的Kimi K3的后训练版本,参数量达3万亿,按计划将于11月4日通过开源许可开放全部模型权重。

第三方AI模型测试机构Artificial Analysis发布的智能指数排名显示,Kimi K3.1 Pro当前测试得分已经超过小米MiMo-V2.6-Pro、DeepSeek V4.1-Flash-1017、Meta Muse Wonder以及OpenAI GPT-6 Astra,待权重正式开放后将成为全球性能排名第一的开源权重AI模型,全品类模型中位列第四。该模型在复杂编码、金融分析、3D创作、软件安全领域表现突出,多步骤工具调用、重复调试迭代类任务得分提升最为明显。Terminal-Bench 4.0测试得分从K3的13%升至67%,超过Claude Sonnet 5.5的64%。CyberGym-E2E-AA测试得分达95%,高于Daybreak Blue的93%。GDPval-AA测试Elo分值达1846,位列全球第三,仅次于Claude Opus 5.5和Muse Wonder。不同测试维度下模型表现存在差异,科学计算类测试中Kimi K3.1 Pro得分仍低于GPT-6 Astra和Claude Opus 5.5, Humanity’s Last Exam、CritPt物理评测等通用推理测试提升幅度相对有限。

现阶段开发者和普通用户可通过Kimi官方网页、移动端应用及API接口访问该模型。标准版API定价为每百万输入token2美元,每百万输出token10美元,每百万缓存输入token0.2美元,价格较Kimi K3下调三分之一,输入输出成本为Claude Opus 5.5和Meta Muse Wonder的一半,但仍高于DeepSeek、小米的同类型模型,DeepSeek在非工作日及工作日非高峰时段还享有额外折扣,价格差距进一步拉大。同步推出的高速版本定价为每百万输入token4美元、输出20美元,同等智能水平下生成速度可达标准版的2至3倍。按完成单个智能指数任务计算,Kimi K3.1 Pro的平均成本约为2.11美元,较Claude Opus 5.5低65%左右。面向开发者的接口文档显示,模型提供low、high、max三档推理力度可选,默认档位为max,对话中切换推理力度会导致前缀缓存失效。模型生成温度固定为1.0,top-p值固定为0.95,支持自动、关闭、强制三种工具调用模式,多轮对话场景下需完整保留模型返回的全部信息包括推理内容才可复现测试效果。企业用户需管理员手动开启该模型访问权限,API用户可选择零数据留存模式。

官方公开的能力测试结果覆盖多类场景。游戏开发领域,模型支持Unreal、Unity、Godot等主流引擎,可结合网页检索、3D生成能力产出完整交互内容。3D创作领域,模型可根据文字需求、手绘草图或参考图生成模型,通过Blender、CAD脚本完成机械、建筑、家具等品类的设计调整,配套3D Make应用支持在线编辑模型、导出3D打印文件及动态展示内容。金融场景下,模型可完成表格建模、投资筛选、盈利分析等工作,内部测试数据显示其金融建模得分较前代提升7.2个点,研报相关能力提升11.6个点,模拟商业管理任务得分达前代的4.2倍。网络安全场景下,模型网络安全指数达64%,可辅助漏洞排查与修复,官方已提前过滤进攻性安全相关训练内容,最终版本漏洞开发测试得分较未过滤版本低9.5个百分点。本次版本性能提升全部来自Kimi K3的大规模后训练流程,覆盖数百个任务域的强化学习环节,训练软件、长上下文处理、算力效率、工具隔离环境等模块均得到优化,未采用全新模型架构。官方披露的行走模拟器实验显示,模型可自主排查仿真环境的计算错误,修正36处肩肌力矩臂计算误差,最终让仿真人体的连续行走时长从最初的数秒提升至11.2秒。

本次产品发布同步引发相关争议。社交平台X上有用户公开疑似Kimi推广活动的指引截图,内容要求参与推广的创作者不得标注付费合作标签,传播中将Kimi包装为对标Claude Opus 5.5的低价开源模型。该用户称自己因上述条款拒绝参与合作。截至10月25日,这份截图尚未得到独立验证,暂无法确认该指引来自月之暗面、外部代理机构还是其他主体,也无法确认是否有创作者按要求发布了未披露商业关系的推广内容,月之暗面暂未对该指控作出公开回应。按照X平台规则及美国联邦贸易委员会相关指引,品牌方与推广者存在付费、免费产品服务等实质性关联时,必须明确披露相关关系。

目前Kimi K3.1 Pro暂不支持用户自托管,有本地部署需求的企业仍需等待11月权重开放。当前API服务部署在新加坡服务器,有数据合规、政府采购要求的用户需提前核验服务条款。此前美国国土安全部曾发布禁令,将DeepSeek排除在民用联邦采购范围之外,要求各机构移除相关产品,该禁令暂未涉及Kimi,但仍提示敏感领域用户提前确认服务合规性。现有测试结果显示,该模型更适合在代码仓库维护、长链路编码、金融文档处理、3D原型制作、漏洞排查修复等场景做定向测试,不建议直接全量替代现有模型,高吞吐量常规任务场景下,DeepSeek、小米等模型仍具备明显价格优势。当前公开测试中Kimi K3.1 Pro幻觉率仍处于较高水平,第三方测试显示其幻觉率从K3的53%降至44%,准确率从48%升至51%,相关输出需经过人工核验,不可直接用于生产环境的自动决策。用户落地选型时需以自身实际工作负载的测试结果为准,综合考量交付成果质量、综合使用成本、人工修正工作量等指标。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

Kimi K3.1 Pro是什么模型?

Kimi K3.1 Pro是月之暗面于2026年10月24日推出的大语言模型,为同年7月发布的Kimi K3的后训练版本,参数量达3万亿,计划于11月4日通过开源许可开放全部模型权重。现阶段开发者和普通用户可通过Kimi官方网页、移动端应用及API接口访问。

Kimi K3.1 Pro适合哪些实际应用场景?

该模型在复杂编码、金融分析、3D创作、软件安全领域表现突出。游戏开发支持Unreal、Unity、Godot等引擎;3D创作可通过Blender、CAD脚本完成设计调整;金融场景可做表格建模、投资筛选、盈利分析;网络安全场景可辅助漏洞排查与修复。建议在代码仓库维护、长链路编码、金融文档处理、3D原型制作等场景做定向测试,不建议直接全量替代现有模型。

Kimi K3.1 Pro与DeepSeek、小米模型相比性能和价格如何?

Artificial Analysis智能指数排名显示,Kimi K3.1 Pro测试得分超过小米MiMo-V2.6-Pro、DeepSeek V4.1-Flash-1017等模型,待权重开放后将成为全球性能排名第一的开源权重AI模型、全品类第四。标准版API每百万输入token2美元、输出10美元、缓存输入0.2美元,较K3降三分之一,为Claude Opus 5.5和Meta Muse Wonder的一半,但仍高于DeepSeek、小米同类型模型。

Kimi K3.1 Pro能本地部署吗?企业接入需要注意什么?

Kimi K3.1 Pro暂不支持用户自托管,有本地部署需求的企业需等待11月4日权重开放。企业用户需管理员手动开启该模型访问权限,API用户可选择零数据留存模式。当前API服务部署在新加坡服务器,有数据合规、政府采购要求的用户需提前核验服务条款;美国国土安全部禁令暂未涉及Kimi,但敏感领域用户仍应提前确认合规性。

Kimi K3.1 Pro的幻觉率和生产可用性怎么样?

第三方测试显示,Kimi K3.1 Pro幻觉率从K3的53%降至44%,准确率从48%升至51%,但仍处于较高水平,相关输出需经过人工核验,不可直接用于生产环境的自动决策。用户落地选型时应以自身实际工作负载的测试结果为准,综合考量交付成果质量、综合使用成本和人工修正工作量。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0