https://www.youngju.dev/blog 천천히 올바르게. AI Researcher & DevOps Engineer Youngju's tech blog. GPU/CUDA, LLM, MLOps, Kubernetes AI workloads, distributed training, and data engineering. ko fjvbn2003@gmail.com (Youngju Kim) fjvbn2003@gmail.com (Youngju Kim) Fri, 15 May 2026 00:00:00 GMT https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive.en https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive.en When DeepSeek-V3 dropped a 671B MoE in December 2024 and R1 added open-weight reasoning in January 2025, the world paused. Since then Alibaba's Qwen 3 (235B-A22B) has become the de facto new open-weight standard; Moonshot's Kimi K2 (1T MoE) owns long-context; Zhipu's GLM-4.5 turned agentic and multimodal; 01.AI's Yi-Large survived a 2025 restructure; ByteDance Doubao 1.5 wins on deployment scale; Tencent Hunyuan T1 joined the reasoning race; Ant Group's Ling/Ming targets finance; StepFun and MiniMax differentiate on multimodal and 4M context. Behind all of this sits a hardware stack rebuilt around US export controls, Huawei Ascend 910C/D, Cambricon, an inference stack of vLLM, LMDeploy, FastGen, and ModelScope, and evaluation benchmarks like SuperCLUE, OpenCompass, and C-Eval. As of May 2026, here is the map of Chinese AI that Korean and Japanese developers actually need to know. Fri, 15 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) aichinadeepseekqwenalibabakimimoonshotglmzhipuyi01aidoubaobytedancehunyuantencentminimaxopen-weight2026deep-diveenglish https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive.ja https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive.ja 2024年12月、DeepSeek-V3が671B MoEを公開し、2025年1月にR1がreasoningまでオープンウェイトで放った瞬間、世界は一瞬止まった。その間にアリババのQwen 3は235B-A22Bで事実上オープンウェイトの新基準となり、MoonshotのKimi K2は1T MoEとロングコンテキストでlong-form王者となり、ZhipuのGLM-4.5はagentic・multimodalへ舵を切り、01.AIのYi-Largeは2025年の再編後も生き残り、ByteDance Doubao 1.5は配備規模で圧倒し、Tencent Hunyuan T1はreasoningへ参戦し、Ant GroupのLing/Mingは金融ドメインを狙い、StepFunとMiniMaxはマルチモーダルと4Mコンテキストで差別化している。その背後には輸出規制とHuawei Ascend 910C/D、Cambricon、vLLM・LMDeploy・FastGen・ModelScopeの推論インフラ、SuperCLUE・OpenCompass・C-Evalの評価エコシステムが敷かれている。2026年5月時点で、韓国・日本のエンジニアが本当に知るべき中国AIの地図を一度に描く。 Fri, 15 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) aichinadeepseekqwenalibabakimimoonshotglmzhipuyi01aidoubaobytedancehunyuantencentminimaxopen-weight2026deep-dive日本語 https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive https://www.youngju.dev/blog/culture/2026-05-15-chinese-ai-labs-2026-deepseek-qwen-kimi-glm-yi-doubao-hunyuan-deep-dive 2024년 12월 DeepSeek-V3가 671B MoE를 공개하고 2025년 1월 R1이 reasoning까지 오픈 가중치로 풀었을 때, 세상은 잠시 멈췄다. 그 사이 알리바바 Qwen 3는 235B-A22B로 사실상 오픈 웨이트의 새 기준이 됐고, Moonshot의 Kimi K2는 1T MoE와 long-context로 long-form 챔피언이 됐고, Zhipu의 GLM-4.5는 agentic·multimodal로 방향을 틀었다. 01.AI의 Yi-Large는 2025년 구조조정 이후에도 살아남았고, ByteDance Doubao 1.5는 배포 규모로 압도하고, Tencent Hunyuan T1은 reasoning에 들어왔고, Ant Group의 Ling/Ming은 금융 도메인을 노리고, StepFun과 MiniMax는 멀티모달과 4M context로 차별화한다. 그 뒤에는 수출 통제와 Huawei Ascend 910C/D, Cambricon으로 굳어진 국산 칩 스택, vLLM·LMDeploy·FastGen·ModelScope의 추론 인프라, SuperCLUE·OpenCompass·C-Eval의 평가 생태계까지 깔려 있다. 2026년 5월 현재, 한국과 일본 개발자가 진짜로 알아야 할 중국 AI 지도를 한 번에 그린다. Fri, 15 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) aichinadeepseekqwenalibabakimimoonshotglmzhipuyi01aidoubaobytedancehunyuantencentminimaxopen-weight2026deep-dive