《Artificial Analysis：2025年第一季度中国人工智能发展状况报告（英文版）（14页）.pdf》由会员分享，可在线阅读，更多相关《Artificial Analysis：2025年第一季度中国人工智能发展状况报告（英文版）（14页）.pdf（14页珍藏版）》请在三个皮匠报告上搜索。

1、State of AI:ChinaArtificial AnalysisQ1 2025Artificial Analysis is a leading and independent AI benchmarking and insights provider.We support engineers and companies to understand AI capabilities and make critical decisions about their AI strategy.Our data,insights and publications are grounded in ou

2、r comprehensive benchmarking of AI technologies and use cases.This includes everything from hourly performance testing of language model APIs to millions of votes in our crowd-sourced arenas.Our public website,artificialanalysis.ai,is widely referenced by companies leading innovation in AI.To discus

3、s this report,our publications,or our services,please get in touch at contactartificialanalysis.ai.1.Artificial Analysis Intelligence Index:average across a range of language model intelligence and reasoning evaluation datasets.Currently includes MMLU,GPQA Diamond,MATH-500&HumanEval.Release date is

4、based on first public launch of the model.2.o3 Intelligence Index estimated by scaling measured Intelligence Index of o1.3.Estimated based on company claims and comparable results where available,not yet independently benchmarked by Artificial AnalysisUS&China:Frontier Language Model Intelligence,Ov

5、er Time1Closing the gap:The final months of 2024 have seen the emergence of the numerous highly performant models from top Chinese AI labs.This has resulted in the delta between the level of intelligence offered by models from Chinese AI labs and US AI labs closing.Several Chinese models are now com

6、petitive with models from the top US labs.Open models close in on the frontier labs:Open weights models,led by those from DeepSeek and Alibaba,have approached o1level intelligence.Reasoning models quickly becoming commonplace:Reasoning models(that“think”before answering)were first introduced by Open

7、AI in 3Q24.Within months,Chinese competitors,led by DeepSeek,have largely replicated the intelligence of o1.Several AI labs in China now have a frontier-level reasoning model.Key TrendsModel Release Date15202530354045505560657075808590951Q232Q233Q234Q231Q242Q243Q244Q241Q254Q222Q25OpenAI,GPT-4OpenAI,

8、GPT-4 TurboOpenAI,GPT-3.5 TurboOpenAI,o1-previewAnthropic,Claude Sonnet(Jun 24)OpenAI,o1OpenAI,o32DeepSeek,V3DeepSeek,R1Alibaba,Qwen 2 Instruct 72BAlibaba,Qwen 2.5 Instruct 72BDeepSeek,V2Alibaba,Qwen Chat 72B3Alibaba,Qwen Chat 7B3GPT-4oChinese AI labs have progressively caught up to US AI labs;model

9、s from Chinese labs are now approaching o1-level intelligence with the release of DeepSeeks R1 modelUSAChinaFRONTIER LANGUAGE MODELS BY ORIGINArtificial Analysis Intelligence Index1OpenAIAnthropicMetaGoogle1.Artificial Analysis Intelligence Index:average across a range of language model intelligence

10、 and reasoning evaluation datasets.Currently includes MMLU,GPQA Diamond,MATH-500&HumanEval.Release date is based on first public launch of the model.2.Estimated based on company claims and comparable results where available,not yet independently benchmarked by Artificial Analysis.3.o3 Intelligence I

11、ndex estimated by scaling measured Intelligence Score of o1.Leading US AI Labs Frontier Language Model Intelligence,Over Time1Competing labs catch up to OpenAIs GPT-4:OpenAI started the language model race in November 2022 with the launch of GPT-3.5 in ChatGPT;leading US labs have largely caught up

12、with frontier models from OpenAI.Big Tech closes in on the frontier labs:Models from Google and Meta are rapidly closing in on frontier models,with Gemini 2.0 Flash exceeding Claude 3.5 Sonnet and GPT 4o capabilities.Sparks of intelligence beyond GPT-4:The final months of 2024 have seen the emergenc

13、e of the first major intelligence leaps beyond GPT-4,led by OpenAIs o3.Topics including reasoning models,data quality and new reinforcement learning techniques have joined pre-training compute scaling as dominant levers for improving models.Key TrendsModel Release Date1520253035404550556065707580859

14、0951Q232Q233Q234Q231Q242Q243Q244Q241Q252Q254Q22GPT-4GPT-4 TurboGPT-4oGPT-3.5 TurboClaude 12Claude 2.12Claude 3OpusClaude 3.5 Sonnet(Jun 24)Claude 3.5 Sonnet(Oct 24)PaLM 2-L2Gemini 1.0 UltraGemini 1.5 Pro(May 24)Gemini 1.5 Pro(Sep 24)Gemini 2.0 FlashLlama 65B2Llama 2 Chat 70B2Llama 3 Instruct 70BLlam

15、a 3.1405BLlama 3.370Bo1o33o1-previewSince the launch of OpenAIs GPT-4 in early 2023,leading US AI labs have scrambled to catch up to OpenAILEADING US FRONTIER LANGUAGE MODELSArtificial Analysis Intelligence Index11.Artificial Analysis Intelligence Index:average across a range of language model intel

16、ligence and reasoning evaluation datasets.Currently includes MMLU,GPQA Diamond,MATH-500&HumanEval.Release date is based on first public launch of the model.2.Estimated based on company claims and comparable results where available,not yet independently benchmarked by Artificial AnalysisLeading Chine

17、se AI Labs Language Model Intelligence,Over Time1Rapid improvements in intelligence:While Chinese AI labs joined the AI race later,they largely closed the intelligence gap with frontier US models in 2024.When OpenAI launched o1,Chinese labs produced a similarly performant model within months(DeepSee

18、ks R1).Leading with open weights models:Chinese AI labs,including Alibaba,DeepSeek and Tencent,have released open weights frontier models that are competitive with the leading models globally.Potential leader in 2025:Early 2025 saw Chinese AI labs,including Alibaba,DeepSeek,MoonShot,Tencent,Zhipu,an

19、d Baichuan prolifically releasing frontier reasoning models.The release velocity and cadence suggest that Chinese AI labs are no longer laggards in 2025.Key TrendsModel Release DateArtificial Analysis Intelligence Index115202530354045505560657075808590951Q232Q233Q234Q231Q242Q243Q244Q244Q222Q251Q25De

20、epSeek V3DeepSeek R1Qwen 2 Instruct 72BQwen 2.5 Instruct 72BDeepSeek V2Qwen 2.5 MaxDeepSeek V12Qwen Chat 72B2Qwen Chat 7B2DeepSeek V2.5Leading Chinese AI labs DeepSeek and Alibaba have steadily released new models,with DeepSeek taking the lead from Alibaba in late 2024DeepSeekAlibabaModel Release Da

21、teLEADING CHINESE FRONTIER LANGUAGE MODELS949089898482828180797978777675747472726455o1,OpenAIR1,DeepSeeko3-mini,OpenAIo1-mini,OpenAIStep-2-16k,StepFunGemini 2.0 Flash(experi-mental),GoogleGemini 1.5 Pro(Sep 24),GoogleClaude 3.5 Sonnet(Oct 24),AnthropicV3,DeepSeekAlibaba Qwen2.5 Max,Alibabao3,OpenAI1

22、Qwen2.5 Instruct 72B,AlibabaMiniMax-Text-01,MiniMaxNova Pro,AmazonLlama 3.3 Instruct 70B,MetaLarge 2(Nov 24),MistralV2.5(Dec 24),DeepSeekGrok Beta,Grok1.5 Large,Jamba,AI21 LabsCommand R+,CohereGPT-4o(Aug 24),OpenAIWhile the US maintains an overall lead in the intelligence frontier,China is no longer

23、 far behind.Few other countries have demonstrated frontier-class trainingThe Language Model Frontier:Country of OriginArtificial Analysis Intelligence Index,Selected Leading Models(Early 2025),Non-exhaustiveLANGUAGE MODEL COUNTRY OF ORIGINFranceCanadaChinaIsraelUSA1.Estimated based on company claims

24、 and comparable results where available,not yet independently benchmarked by Artificial Analysis2.A number of leading models from Chinese AI labs are excluded due to limited access or evaluation data9489878483828281807979787776747373706552R1,DeepSeekKimi k1.5,Moonshot1Step-R-mini,StepFun1M1-Preview,

25、Baichuan1Step-2-16k,StepFunGemini 2.0 Flash Experi-mental,GoogleGLM-Zero-Preview,Zhipu1Doubao 1.5 Pro,Bytedance1V3,DeepSeeko3,OpenAIQwQ,AlibabaDoubao 1.5 Lite,Bytedance1MiniMax-Text-01,MiniMaxHunyuan Large,Tencent1Ernie 4.0 Turbo,Baidu1Yi-Lightning,Yi AI1GLM-4-Plus,Zhipu14-Turbo,Baichuan1V1-128k,Moo

26、nshot1Qwen 2.5 Max,AlibabaAs of early 2025,several Chinese AI labs have demonstrated or claimed frontier-level intelligence,with seven releasing models featuring reasoning capabilitiesThe Language Model Frontier:Models by Chinese AI LabsArtificial Analysis Intelligence Index,Leading Models(Early 202

27、5),Non-exhaustive1.Estimated based on company claims and comparable results where available,not yet independently benchmarked by Artificial AnalysisLANGUAGE MODEL COUNTRY OF ORIGINHighest Intelligence US Reasoning ModelHighest Intelligence US Non-reasoning ModelThe leading Chinese Big Tech firms are

28、 actively competing in the AI race and have released AI language models as well as models across other modalities1.Market cap as per Reuters(aa 31 Jan 25)2.ByteDance is a private company.Valuation by Reuters 3.Huawei is a private company.Valuation by Reuters(2023)4.Artificial Analysis Intelligence I

29、ndex:average across a range of language model intelligence and reasoning evaluation datasets.Currently includes MMLU,GPQA Diamond,MATH-500&HumanEval.5.Estimated based on company claims and comparable results where available,not yet independently benchmarked by Artificial AnalysisFrontier Models by C

30、hinese Big Tech FirmsCHINA AI LABS OVERVIEW:BIG TECHAlibabaBaiduByteDanceHuaweiTencentDescriptionLarge ecommerce player and Hyperscaler(Alibaba Cloud),largest shareholder of Ant GroupChinas largest search engine,and operator of Wenxin Yiyan,an AI chatbot with a reported 300m usersParent company of D

31、ouyin(TikTok)and Toutiao,one of Chinas leading news applicationsGlobal telco leader and one of the worlds largest smartphone manufacturersParent company of Riot Games and WeChat,the all-in-one super app of China;Hyperscaler with their Tencent Cloud offeringAI Strategy(high-level)Release open weights

32、 modelsMore recently launched proprietary modelsOffer inference on Alibaba CloudActively integrating proprietary models into search platformLong time leader in self-driving AIDevelop proprietary models and integrate across their consumer platformsDevelop proprietary,domain-specific models and offer

33、on Huawei CloudRelease open weights models and offer proprietary models on Tencent CloudBest LLM4Non-ReasoningQwen 2.5 MaxIntelligence:79Ernie 4.0 TurboIntelligence:765Doubao 1.5 LiteIntelligence:775Pangu 5.0 LargeHunyuan LargeIntelligence:74ReasoningQwQIntelligence:785Doubao 1.5 ProIntelligence:805

34、Other ModelsText to SpeechSpeech to SpeechImage GenerationVideo Generation3D GenerationPrimary Consumer AppsTongyi QianwenWenxin Yiyan,Wenxin YigeDoubaoCeliaYuanbao,YuanqiValuation(US$)235B132B1300B2128B3469B1Non-ExhaustiveOpen Weights LLMOther Firms with AI AmbitionsChinas largest provider of Inter

35、net and mobile security products.Launched the Zhinao series of models under the 360 AI brandBeijing-based internet group with 300m MAUs;owner of the Opera browser.Launched the SkyWork series of models and AI acceleratorsLeading voice AI company in China with 14,000 employees.Launched the Spark serie

36、s of modelsKunlun TechSHE:300418(Mkt Cap:$6B)1360 Security(Qihoo 360)SHA:601360(Mkt Cap:$11B)1iFlytekSHE:002230(Mkt Cap:$16B)1MeituanHKG:3690(Mkt Cap:$115B)1Chinas leading shopping platform with 600m DAUs.Cofounder Wang Huiwen returned to lead AI efforts.Investor in multiple frontier AI labsChinas l

37、eading consumer electronics brand.Launched the MiLM series of small models.Recently poached Luo Fuli,DeepSeek researcher,to run AI lab.Investor in multiple frontier AI labsXiaomiHKG:1810(Mkt Cap:$123B)1Chinese AI startups,with the support of Chinese Big Tech firms and the Chinese Government,have dev

38、eloped some of the worlds leading open weights models1.Artificial Analysis Intelligence Index:average across a range of language model intelligence and reasoning evaluation datasets.Currently includes MMLU,GPQA Diamond,MATH-500&HumanEval.2.Estimated based on company claims and comparable results whe

39、re available,not yet independently benchmarked by Artificial Analysis 2.Pitchbook(Mar 2024)3.Pitchbook(Aug 2024)4.Pitchbook(Dec 24)5.Pitchbook(Jul 24)6.Pitchbook(Aug 24)Frontier Models by Chinese AI Tigers and StartupsMiniMaxMoonShot AI01 AIDeepSeekZhipuBaichuanStepfunDescriptionChina AI Tiger and p

40、ublisher of Talkie AI app(4th most downloaded in US in 1H24)China AI Tiger with 2M Chinese character context window model;Chinas most well-funded AI startup based on available informationChinese AI startup focused on smaller language models founded by Lee Kai-Fu(author,former head of Google China)Ch

41、inese AI lab originating out of an AI-focused quantitative trading firmChina AI Tiger with nearly 700k enterprise and developer usersChina AI Tiger with a focus on medical AI models founded by Wang Xiaochuan(ex-CEO,Sogou)First Chinese AI startup to develop a trillion-parameter model;founded by Jiang

42、 Daxin(ex-Chief Scientist,Microsoft Research Asia)Best LLM1Non-ReasoningMiniMax-Text-01 Intelligence:76V1-128kIntelligence:52Yi-LightningIntelligence:73V3Intelligence:79GLM-4-PlusIntelligence:70Baichuan 4-TurboIntelligence:65Step-2-16kIntelligence:82ReasoningKimi k1.5Intelligence:87R1Intelligence:89

43、GLM-Zero-PreviewIntelligence:81Baichuan M1-PreviewIntelligence:83Step-R-miniIntelligence:84Other ModelsText to SpeechSpeech to SpeechImage GenerationVideo Generation3D GenerationPrimary Consumer AppsHailuo AI Chat,Hailuo AI VideoKimiYiChatDeepSeek ChatChatGLMBai XiaoyingYuewen,PopDuckFunding Raised(

44、$)0.85B21.67B30.2B4Unknown1.12B51.04B6UnknownState Backed EntityOpen Weights LLMNon-ExhaustiveNetworkIconStepfunCHINA AI LABS OVERVIEW:TIGERS&STARTUPSNotable Investors(non-exhaustive)Escalating regulatory restrictions have banned the export of high-end AI accelerators to China(1/2)Commentary NVIDIA

45、reacted quickly to both the October 2022 and October 2023 controls by releasing Hopper GPU variants that complied/comply with the regulations.Specifically,after the H100 and A100 were banned for export to China,NVIDIA released the H800 and A800 with limited interconnect(see appendix for full Hopper

46、generation specifications).The October 2023 controls went on to ban export of the H800 and A800 to China,leading to NVIDIA developing the H20 to continue selling a Hopper-generation GPU to Chinese customers.The H20 has limited compute(148 TFLOPs)compared to the H100(989 TFLOPs)Regulatory Restriction

47、sNVIDIA GPU ArchitectureModelPre-ControlsOctober 2022 Controls2October 2023 Controls3,4AI Diffusion Rules5Announced7-Oct-2217-Oct-2313-Jan-25Effective121-Oct-2217-Nov-2315-May-25BlackwellB200B100HopperH100H200H800H20LovelaceL40SL4L40L20L2AmpereA100A800A40A30Consumer GPUsRTX 6000 AdaRTX 4090RTX 4090D

48、RTX 30901.Effective date refers to latest compliance date 2.BIS 3.Georgetown CSET 4.Federal Register 5.BISNo Licence RequiredUnreleasedNAC License RequiredPresumption of DenialEXPORT RESTRICTIONS TIMELINEEscalating regulatory restrictions have banned the export of high-end AI accelerators to China(2

49、/2)1.TPP measured in Tera Operations per Second,PD measured as TPP/Die Size.2.Effective date refers to latest compliance date 3.BIS 4.Georgetown CSET 5.Federal Register 6.BIS 7.Federal Register 8.BIS 9.Federal RegisterRuleSummaryDates1Impact2October 2022 Controls3Initial restrictions on frontier GPU

50、s.Both performance and interconnect thresholds had to be breached for the GPU to be restricted.Announced:7-Oct-22Restriction ClassificationCriterionTotal Processing Performance(TPP)TPP 4,800Effective:21-Oct-22Interconnect BandwidthTPP 600 GB/sOctober 2023 Controls4,5Revised framework to prevent work

51、arounds.Restricted exports of GPUs to China based on TPP or Performance Density(PD)Announced:17-Oct-23GroupingsCriterion(Datacenter GPUs)Group 1:Presumption of denialTPP 4,800 or TPP 1600 AND PD 5.92.Effective:17-Nov-23Group 2:Restrictive NAC licensing review2,400 TPP 4,800 AND PD 1.6or TPP 1,600 AN

52、D PD 3.2.Group 3:No RestrictionsTPP 1,600 or PD 3.2BIS Final Rule6Crackdown on indirect imports by Chinese-affiliated chip manufacturing entitiesAnnounced:2-Dec-24Did not impact restricted chips140 entities(majority Chinese)from advanced chip sector now face a presumption of denial and added to Enti

53、ty List in Dec 247Effective:31-Dec-24Updated:16-Jan-24AI Diffusion Rule8Extensive three-tiered licensing framework segregating access to GPUs by countriesAnnounced:13-Jan-25 Tier 3 countries(including China)face ade facto banon advanced AI chipsAll exports of controlled chips to these Tier 3 countri

54、es now require an export license,subject to apresumption of denialduring reviewTier 2 countries now face limitations on large orders of AI chipsEffective:15-May-25AI Due Diligence Rule9Companion KYC rule for AI Diffusion RuleAnnounced:16-Jan-25 Requires companies to conduct KYC-like compliance check

55、s on their customers and comply with the AI Diffusion RuleEffective:31-Jan-24 on-ai-technologiesRegulatory RestrictionsEXPORT RESTRICTIONS TIMELINEUS export controls restrict export of leading Nvidia accelerators based on performance and density thresholds;the H20 and L20 fall below these thresholds

56、 and can be freely exportedCommentary The H20 and L20 are the only current NVIDIA data center-class AI accelerators that do not exceed either the Total Processing Performance or Performance Density threshold.While the H20 accelerator is currently available for sale in China,the Trump administration

57、has started preliminary conversations around the potential inclusion of the chip on the restricted list,suggesting that there may be a further broadening of the scope of restricted chipsUS Accelerators Prohibited for Export to China1,205,00052040,0002530354045025,00010,00020,00015,000Total Processin

58、g Performance(TOPS)3L4MI300XB200B100Performance DensityH100H200L40SA800A100L20L40H20AMDNVIDIAPresumption of DenialNo Licence RequiredNAC License Required1.SemiAnalysis 2.Georgetown CSET 3.Total Processing Performance(TPP)measured in Tera Operations per Second,Performance Density measured as TPP/Die

59、SizeArtificial Analysishelloartificialanalysis.aihttps:/artificialanalysis.ai/Legal notice:Copyright 2025 Artificial Analysis,Inc.All rights reserved.This document,including any data,analysis,and insights contained herein,is provided by Artificial Analysis for informational purposes only.The informa

60、tion is based on data collected through various sources,including but not limited to first party benchmarking and surveys conducted on our website.While Artificial Analysis strives to ensure the accuracy and reliability of the information,it is provided“as is”and may not be complete or up to date.Th

61、e content should not be construed as professional advice,and recipients are encouraged to conduct their own research and analysis before making any decisions based on this information.By accessing or using this document,you agree to be bound by Artificial Analysiss Terms of Service,available on our

62、website.Appendix:Accelerator hardware specifications(NVIDIA Hopper,NVIDIA Blackwell,AMD)NVIDIA H100(SXM)NVIDIA H100(NVL)NVIDIA H100(PCIe)NVIDIA H800(PCIe)NVIDIA HGX H20NVIDIA H200(NVL)NVIDIA H200(SXM)NVIDIA B200NVIDIA GB2001AMD MI300XAMD MI325XInitial Release Date1Q231Q231Q232Q234Q232Q242Q241Q251Q25

63、4Q234Q24Memory80GB HBM394GB HBM380GB HBM2e80GB HBM2e96GB HBM3141GB HBM3e141GB HBM3e192GB HBM3e384GB HBM3e192GB HBM3256MB on-chip SRAM256GB HBM3e256MB on-chip SRAMMemory Bandwidth3.35 TB/s3.9 TB/s2 TB/s2 TB/s4 TB/s4.8 TB/s4.8 TB/s8 TB/s16 TB/s5.3 TB/s6 TB/sPower/TDP700W350-400W350W350W400W600W700W1,0

64、00W2,700W750W1000WBF/FP16 TFLOPs(Dense)989 TFLOPs835 TFLOPs756 TFLOPs756 TFLOPs148 TFLOPs835 TFLOPs989 TFLOPs2,250 TFLOPs5,000 TFLOPs1,307 TFLOPs1,307 TFLOPsChip-to-chip Interconnect900GB/s NVLink600GB/s NVLink600GB/s NVLink400GB/s NVLink900GB/s NVLink900GB/s NVLink900GB/s NVLink1,800 TB/s NVLink3,600GB/s NVLink7X128GB/s Infinity Fabric7X128GB/s Infinity FabricModule TypeSXMPCIePCIePCIeSXMPCIeSXMSXMSXMProcess NodeTSMC 4NTSMC 4NTSMC 4NTSMC 4NTSMC 4NTSMC 4NTSMC 4NTSMC 4NPTSMC 4NPTSMC 5NTSMC 5NSource URLhttps:/ Blackwell superchip includes two Blackwell GPUs and a NVIDIA Grace ARM CPU