The　computational　efficiencies　of　traditional　reliability　methods,　such　as　Monte　Carlo　(MC),　are　extremely　low.　There　are　also　some　shortcomings　for　surrogate　model　(SM)-based　methods,　e.g.,　the　sample　points　of　the　quadratic　polynomial　(QP)-MC　grow　exponentially　with　the　increases　of　random　variables　and　the　artificial　neural　network　(ANN)-MC　may　exhibit　overfitting　with　limited　sample　numbers,　etc.　However,　the　characteristic　of　support　vector　machine　(SVM)　is　that　it　specifically　fits　for　small　samples　and　has　strong　learning　and　good　generalization　abilities　so　that　it　can　obtain　an　optimal　solution　even　with　limited　samples.　In　this　case,　a　high-efficiency　and　high-accuracy　dynamic　reliability　framework　called　as　SVM-based　classification　extremum　method　(SVM-CEM)　combining　SVM　classification　theory　with　random　probability　model　based　on　optimization　idea　is　proposed,　which　is　very　suitable　for　the　flexible　mechanism　(FM)　that　has　few　samples.　First,　an　implicit　limit　state　equation　(LSE)　of　dynamic　response　and　a　reliability　model　with　multiple　failure　modes　for　FM　are　established.　The　kernel　function　is　introduced　in　building　the　model,　the　solution　of　optimal　classification　hyperplane　is　translated　into　a　dual　problem　of　convex　quadratic　programming　optimization,　which　is　regarded　as　the　surrogate　model　of　FM＇s　dynamic　response　extreme　value　(DREV).　Then,　this　method　is　used　to　analyze　the　dynamic　reliability　of　FM＇s　maximum　angular　acceleration　(MAA).　Finally,　to　reveal　the　validity　of　this　method,　SVM-CEM　is　compared　with　MC,　QP-MC,　and　ANN-MC.　The　conclusion　is　that　the　computational　efficiency　of　SVM-CEM　is　better　than　that　of　MC,　QP-MC,　and　ANN-MC　ensuring　the　computational　accuracy.　The　proposed　SVM-CEM　in　dynamic　reliability　analysis　has　important　guiding　significance　for　the　application　of　FM＇s　practical　engineering.