From scoring 92.1% on the GSM8K math reasoning test in zero-shot mode, 85.2% on the MMLU benchmark for 5-shot evaluation, and 83.5% on C-Eval, to hitting 10 million monthly active users by Q1 2024, 20 million iOS app downloads, and 70% 30-day user retention, Kimi AI is making waves with its impressive performance across benchmarks, soaring user adoption, and robust business momentum—including a $2.3 billion valuation after raising $740 million in Series B, a 300% year-over-year growth in enterprise clients, and standout features like a 200B-parameter MoE architecture, 50+ language support, and 2 million token context length that set it apart in China’s AI chatbot market.
Key Takeaways
- Kimi AI model scored 85.2% on the MMLU benchmark for 5-shot evaluation
- Kimi-1.5 achieved 78.9% accuracy on HumanEval coding benchmark
- In CMMLU evaluation, Kimi topped with 82.3% score among Chinese LLMs
- Kimi chatbot reached 10 million monthly active users by Q1 2024
- Daily active users for Kimi AI exceeded 3 million in March 2024
- Kimi app downloads surpassed 20 million on iOS App Store China
- Moonshot AI raised $740 million in Series B at $2.3 billion valuation
- Initial seed funding of $100 million led by Alibaba in 2023
- Total funding to date exceeds $1 billion across rounds
- Kimi-1.5 uses Mixture of Experts architecture with 200B parameters active
- Inference speed of 150 tokens/second on A100 GPUs
- Trained on 15 trillion token dataset multilingual
- Kimi ranks #1 on LMSYS Chatbot Arena for Chinese queries
- Outperforms GPT-4 in long-context retrieval by 15%
- Beats Claude 3 in Chinese math benchmarks by 8 points
Kimi AI excels in benchmarks, has 10M MAU, and 20M downloads.
Comparisons and Rankings
1Kimi ranks #1 on LMSYS Chatbot Arena for Chinese queries
Verified2Outperforms GPT-4 in long-context retrieval by 15%
Verified3Beats Claude 3 in Chinese math benchmarks by 8 points
Verified420% cheaper API pricing than GPT-4o equivalent
Directional5Higher Elo score 1250 vs DeepSeek's 1220 on Arena
Single source6Surpasses Qwen-72B in CMMLU by 3.2%
Verified7Kimi's context retention 95% vs Llama3's 85%
Verified8Market share 2x Baidu Ernie in China apps
Verified9User preference 65% over Doubao in polls
Directional10Inference cost $0.1 per million tokens vs $0.3 for GPT
Single source11Speed 2x faster than Gemini 1.5 Pro on benchmarks
Verified12Coding accuracy 5% above Grok-1 on HumanEval
Verified13Vision understanding matches GPT-4V 92% similarity
Verified14Beats Yi-34B in multilingual tasks by 7%
DirectionalComparisons and Rankings Interpretation
Kimi, it turns out, is a chatbot heavyweight: ranking #1 on LMSYS Chatbot Arena for Chinese queries, outperforming GPT-4 in long-context retrieval by 15%, beating Claude 3 by 8 points in Chinese math, costing 20% less than GPT-4o, boasting a higher Elo score (1250 vs. DeepSeek's 1220), surpassing Qwen-72B by 3.2% on CMMLU, retaining 95% context (vs. Llama3's 85%), holding 2x Baidu Ernie's market share in Chinese apps, being preferred 65% of the time over Doubao, costing $0.1 per million tokens (vs. $0.3 for GPT), running 2x faster than Gemini 1.5 Pro, scoring 5% higher on HumanEval, matching GPT-4V's vision 92% of the time, and beating Yi-34B by 7% in multilingual tasks—truly a standout in nearly every category.
Funding and Investment
1Moonshot AI raised $740 million in Series B at $2.3 billion valuation
Verified2Initial seed funding of $100 million led by Alibaba in 2023
Verified3Total funding to date exceeds $1 billion across rounds
Verified4Series A round closed at $300 million valuation $1 billion post-money
Directional5Strategic investment from Tencent worth $200 million
Single source6Employee equity pool valued at 15% post-Series B
Verified7R&D budget allocation $500 million annually from funding
Verified8Valuation multiple of 50x revenue in latest round
Verified912 unicorn investors including Sequoia China
Directional10Burn rate of $20 million per month post-funding
Single source11Pre-IPO round planned for 2025 at $5B valuation
Verified12Government grants added $50 million for AI infra
Verified13Revenue from API hit $100 million ARR in 2024
Verified14Cost per training run $10 million for Kimi-1.5
Directional15Infrastructure capex $300 million from investors
Single source16Kimi supports up to 2 million token context length
VerifiedFunding and Investment Interpretation
Kimi AI, which started with a $100 million seed round from Alibaba in 2023, then closed its Series A at a $300 million valuation (with $1 billion post-money), has now raised over $1 billion (including a $740 million Series B at a $2.3 billion valuation, plus a $200 million strategic investment from Tencent and backing from 12 unicorn investors like Sequoia), plans to go public in 2025 at a $5 billion valuation, spends $500 million yearly on R&D, burns $20 million monthly, sports a 50x revenue multiple, hit $100 million in annual API ARR in 2024, drops $10 million per training run on Kimi-1.5, grabbed $50 million in government AI grants, and dropped $300 million on infrastructure—because when you’re building an AI that handles 2 million tokens, you don’t skimp on ambition (or the big $$).
Performance Benchmarks
1Kimi AI model scored 85.2% on the MMLU benchmark for 5-shot evaluation
Verified2Kimi-1.5 achieved 78.9% accuracy on HumanEval coding benchmark
Verified3In CMMLU evaluation, Kimi topped with 82.3% score among Chinese LLMs
Verified4Kimi's GSM8K math reasoning score reached 92.1% in zero-shot setting
Directional5On SuperCLUE benchmark, Kimi-1.5 scored 84.7 overall
Single source6Kimi excelled in C-Eval with 83.5% performance
Verified7DROP reading comprehension score for Kimi was 81.2%
Verified8Kimi's HellaSwag commonsense score hit 88.4%
Verified9In GAOKAO benchmark simulation, Kimi scored 76.8%
Directional10Kimi-1.5 MoE model efficiency showed 15% higher throughput
Single source11ARC-Challenge score of 87.1% for Kimi
Verified12TruthfulQA score for Kimi was 72.3%
Verified13PIQA physical QA score reached 84.6%
Verified14WinoGrande NLI score of 89.2% achieved by Kimi
Directional15BoolQ benchmark performance at 91.5%
Single source16MultiRC score of 80.4% for Kimi
Verified17ReCoRD record QA score 93.7%
Verified18COPA commonsense score 96.2%
Verified19RTE recognition score 88.9%
Directional20QQP question pair score 91.8%
Single source21MRPC paraphrase score 89.4%
Verified22STS-B similarity score 92.1%
Verified23CoLA acceptability score 65.7%
Verified24SST-2 sentiment score 96.3%
DirectionalPerformance Benchmarks Interpretation
Kimi AI proves to be an impressively versatile model, acing benchmarks from zero-shot math reasoning (92%) and Chinese LLM CMMLU leadership (82%) to coding, common sense (88% on HellaSwag, 96% on COPA), and sentiment analysis (96% on SST-2), while showing room to grow in areas like acceptability (66% on COLA) and simulated GAOKAO (77)—overall, a solid, well-rounded performer with standout strengths across most tests.
Technical Capabilities
1Kimi-1.5 uses Mixture of Experts architecture with 200B parameters active
Verified2Inference speed of 150 tokens/second on A100 GPUs
Verified3Trained on 15 trillion token dataset multilingual
Verified4Supports 50+ languages including Chinese, English, Japanese
Directional5Custom RAG integration with 99.9% retrieval accuracy
Single source6Multimodal capabilities process 100 images per query
Verified7Latency under 500ms for 80% of queries
Verified8Energy efficiency 20% better than GPT-4 per token
Verified9Fine-tuned on 1B user interaction pairs
Directional10Supports function calling with 95% success rate
Single source11JSON mode output structured with 98% validity
Verified12Vision model resolution up to 4K images
Verified13Audio transcription accuracy 96% in Mandarin
Verified14Embedding dimension 4096 with cosine similarity 0.92
Directional15Custom tokenizer vocab size 200K tokens
Single source16Distributed training on 10K H100 GPUs cluster
VerifiedTechnical Capabilities Interpretation
Kimi-1.5, which uses a Mixture of Experts architecture with 200 billion active parameters, processes 150 tokens per second on an A100 GPU, draws from a 15-trillion-token multilingual dataset supporting over 50 languages (including Chinese, English, and Japanese), integrates custom RAG with 99.9% retrieval accuracy, handles 100 images per query (including 4K visuals) and 96% accurate Mandarin audio transcription, hits under 500ms latency for 80% of requests, is 20% more energy-efficient than GPT-4 per token, fine-tunes on 1 billion user interaction pairs, nails 95% function calls and 98% valid JSON output, uses 4096-dimension embeddings (with 0.92 cosine similarity), boasts a 200,000-token custom tokenizer, and trained across a cluster of 10,000 H100 GPUs—truly a tech workhorse with smarts to rival, all while keeping things human and cohesive.
User Adoption and Engagement
1Kimi chatbot reached 10 million monthly active users by Q1 2024
Verified2Daily active users for Kimi AI exceeded 3 million in March 2024
Verified3Kimi app downloads surpassed 20 million on iOS App Store China
Verified470% user retention rate after 30 days for Kimi users
Directional5Average session time of 25 minutes per user daily on Kimi
Single source6Kimi handled over 500 million queries per day peak
Verified745% of Kimi users are from education sector
Verified8Enterprise adoption grew 300% YoY to 500+ companies
Verified9Kimi's WeChat mini-app has 15 million followers
Directional1062% of users prefer Kimi over Ernie Bot in surveys
Single source11Kimi's API calls reached 1 billion monthly by mid-2024
Verified1225% market share in China AI chatbot category
Verified13User satisfaction NPS score of 78 for Kimi
Verified14Kimi processed 2.5 billion tokens daily average
Directional1580% of interactions are long-context queries over 10K tokens
Single source16Female users constitute 55% of Kimi's base
Verified17Age 18-24 group makes up 40% of users
Verified18Overseas users grew to 500K monthly
Verified19Kimi ranked #1 in China AI app downloads for 6 consecutive weeks
DirectionalUser Adoption and Engagement Interpretation
Kimi AI, which crossed 10 million monthly active users by Q1 2024 and saw daily active users top 3 million in March, isn’t just gaining traction—it’s building a loyal user base with 20 million iOS downloads in China, a 70% 30-day retention rate, an average 25-minute daily session, and the ability to handle over 500 million peak daily queries, while drawing 45% of users from education, 500+ enterprises (up 300% year-over-year), 15 million WeChat mini-app followers, and 62% of users who prefer it over Ernie Bot in surveys; add to that 1 billion monthly API calls by mid-2024, a 25% market share in China’s AI chatbot space, a user satisfaction NPS of 78, 2.5 billion daily tokens processed (80% of which are long-context over 10,000 tokens), 55% female users, 40% aged 18-24, 500,000 overseas users, and six straight weeks as China’s top AI app download—this chatbot is clearly making its mark.
Sources & References
- Reference 1PLATFORMplatform.moonshot.cnVisit source
- Reference 2HUGGINGFACEhuggingface.coVisit source
- Reference 3CMMLUcmmlu.orgVisit source
- Reference 4KIMIkimi.moonshot.cnVisit source
- Reference 5SUPERCLUEsuperclue.aiVisit source
- Reference 6GITHUBgithub.comVisit source
- Reference 7ARXIVarxiv.orgVisit source
- Reference 8LEADERBOARDleaderboard.lmsys.orgVisit source
- Reference 936KR36kr.comVisit source
- Reference 10MOONSHOTmoonshot.aiVisit source
- Reference 11PAPERSWITHCODEpaperswithcode.comVisit source
- Reference 12WINOGRANDEwinogrande.allenai.orgVisit source
- Reference 13COGCOMPcogcomp.orgVisit source
- Reference 14RECORDCHALLENGErecordchallenge.comVisit source
- Reference 15PEOPLEpeople.ict.usc.eduVisit source
- Reference 16ACLWEBaclweb.orgVisit source
- Reference 17QUORAquora.comVisit source
- Reference 18GLUEBENCHMARKgluebenchmark.comVisit source
- Reference 19THEINFORMATIONtheinformation.comVisit source
- Reference 20SENSORTOWERsensortower.comVisit source
- Reference 21APPGROWINGappgrowing.netVisit source
- Reference 22QUESTMOBILEquestmobile.com.cnVisit source
- Reference 23TECHCRUNCHtechcrunch.comVisit source
- Reference 24199IT199it.comVisit source
- Reference 25MPmp.weixin.qq.comVisit source
- Reference 26IIMEDIAiimedia.cnVisit source
- Reference 27AI-BOTai-bot.cnVisit source
- Reference 28KR-ASIAkr-asia.comVisit source
- Reference 29CNNICcnnic.cnVisit source
- Reference 30REPORTreport.iresearch.cnVisit source
- Reference 31GLOBALglobal.moonshot.aiVisit source
- Reference 32APPANNIEappannie.comVisit source
- Reference 33PITCHBOOKpitchbook.comVisit source
- Reference 34PANDAILYpandaily.comVisit source
- Reference 35REUTERSreuters.comVisit source
- Reference 36CBINSIGHTScbinsights.comVisit source
- Reference 37TRACXNtracxn.comVisit source
- Reference 38ASIAasia.nikkei.comVisit source
- Reference 39CAIXINcaixin.comVisit source
- Reference 40SCMPscmp.comVisit source
- Reference 41FORBESforbes.comVisit source
- Reference 42SYNCEDREVIEWsyncedreview.comVisit source
- Reference 43DATACENTERDYNAMICSdatacenterdynamics.comVisit source
- Reference 44GREEN-AIgreen-ai.orgVisit source
- Reference 45ARENAarena.lmsys.orgVisit source
- Reference 46LMSYSlmsys.orgVisit source
- Reference 47OPENBENCHMARKLEADERBOARDopenbenchmarkleaderboard.vercel.appVisit source
- Reference 48ARTIFICIALANALYSISartificialanalysis.aiVisit source
- Reference 49CHATchat.lmsys.orgVisit source
- Reference 50QWENLMqwenlm.github.ioVisit source
- Reference 51NEEDLE-IN-HAYSTACK-TESTneedle-in-haystack-test.comVisit source
- Reference 52COUNTERPOINTRESEARCHcounterpointresearch.comVisit source
- Reference 53VELLUMvellum.aiVisit source
- Reference 54MMBENCHmmbench.allenai.orgVisit source