Spectral　Reconstruction　(SR)　has　attracted　a　lot　of　attention　in　recent　years.　A　few　studies　have　proposed　spectral　reconstruction　techniques　based　on　Transformers.　However,　the　performance　optimization　of　Transformers　still　is　a　vital　issue.　Furthermore,　how　to　simultaneously　explore　the　spatial　and　spectral　information　of　hyperspectral　image　and　fuse　the　features　of　different　resolutions　at　different　hierarchical　levels　is　worth　further　investigation.　Considering　these　issues,　in　this　paper,　we　propose　the　Patch　Clustering-Based　Spectral-　Transformer　with　Multi-level　Fusion　Network　(PCFM).　Specifically,　to　address　the　interaction　of　pixels　with　similar　long-distance　spectral　information,　a　Patch-based　Clustering　Spectral　Self-Attention　(PCSS)　mechanism　is　developed.　After　patching,　each　patch　can　focus　on　capturing　local　features,　and　clustering　can　reduce　the　computational　complexity　of　processing　large　images.　Furthermore,　in　order　to　fully　consider　the　relationship　between　different　clusters　for　PCSS,　a　Deal　Conv　Feed　Attention　(DCFA)　module　is　designed,　which　adjusts　the　weight　of　the　input　data　by　calculating　the　correlation　between　clusters　so　that　the　model　can　better　understand　and　handle　these　complex　interactions.　In　addition,　to　address　the　various　scales　fusion　issue,　different　from　the　previous　methods,　an　Hierarchical　Optimization　Feature　Fusion　mechanism　(HOFF)　is　introduced　for　merging　images　generated　at　different　hierarchical　levels.　Finally,　the　Channel　Relation　Module　(CRM)　is　employed　to　further　bridge　the　gap　between　the　reconstructed　HSI　and　their　ground　truth.　Extensive　experimental　results　on　the　Pavia,　Sandiga　and　Chikusei　datasets　indicate　that　this　method　outperforms　the　SOTA　method　in　quantitative　metrics　and　visual　effects.