The　most　current　Chinese　named　entity　recognition　models　can　be　divided　into　the　following　two　types:　one　is　based　on　character　unit　and　the　other　is　based　on　word　unit.　Both　of　them　have　their　disadvantages.　The　former　lacks　semantic　and　boundary　information　of　words,　while　the　latter　depends　on　the　accuracy　of　word　segmentation.　In　addition,　Chinese　characters　evolved　from　hieroglyphics,　which　contain　rich　semantic　information.　To　address　the　above　problems,　this　paper　proposes　a　Chinese　named　entity　recognition　model　with　the　multi-granularity　embedding　from　Chinese　strokes,　characters,　and　words.　The　method　first　obtains　the　stroke　list　corresponding　to　Chinese　characters　in　the　training　corpus,　and　uses　a　convolutional　kernel　with　multi-window　size　to　obtain　the　n-gram　information　of　strokes.　Secondly,　a　convolutional　neural　network　is　used　to　obtain　the　local　information　of　the　sequence　and　the　representation　of　Chinese　characters　within　the　local　area.　Finally,　considering　the　constraint　relationship　between　the　information　of　the　context　and　the　label,　the　BiLSTM　network　is　used　for　context　modeling,　and　the　CRF　network　is　used　as　the　label　decoding　layer.　The　experimental　results　show　that　the　F1-Score　of　this　method　achieves　62.41%　and　95.28%　on　Weibo　and　MSRA　datasets,　respectively.　©　2022　IEEE.