Steel　fiber　is　usually　used　to　improve　ultra-high-performance　fiber-reinforced　concrete　(UHPFRC)　tensile　properties　and　remedy　its　brittle　failure.　X-ray　CT　scanning　(X-CT)　was　used　to　observe　the　orientation　and　distribution　of　steel　fibers　inside　UHPFRC.　It　revealed　that　the　steel　fiber　inside　UHPFRC　distribution　angle　is　not　equal　probability,　and　the　probability　is　higher　along　the　pouring　direction　in　uniaxial　tensile　specimen　mold.　The　uniaxial　tensile　tests　(UTT)　of　19　UHPFRC　specimens　with　different　fiber　volume　fractions　and　aspect　ratios　were　performed　to　analyze　the　steel　fiber　on　the　tensile　response　and　tension　failure　patterns　of　UHPFRC.　Test　results　indicated　that　with　an　increase　in　the　volume　fraction　or　aspect　ratio　of　steel　fiber,　the　first-cracking　strength,　tensile　strength　and　peak　tensile　strain　of　UHPFRC　increase;　while　they　have　different　growth　rates.　In　addition,　with　the　increase　of　volume　fraction　and　aspect　ratio,　the　failure　patterns　of　UHPFRC　tension　specimens　gradually　changed　from　brittle　failure　with　single　crack　to　ductile　failure　with　multiple　cracks.　The　stress　superposition　compositions　of　tensile　strength　were　different　under　different　failure　patterns.　Combining　mechanical　analysis　and　numerical　statistical　analysis,　prediction　equations　of　the　tensile　strength　of　strain　softening　UHPFRC　and　pseudo-strain　hardening　UHPFRC　considering　the　effect　steel　fiber　parameters　were　respectively　proposed.　Through　the　mesoscopic　equivalent　model　analysis　of　UHPFRC　uniaxial　tensile　specimens,　the　functional　relationships　between　the　effect　of　steel　fiber　parameters　on　the　first-cracking　strength　and　peak　tensile　strain　were　analyzed,　and　then　their　prediction　equations　were　proposed,　combined　with　numerical　statistical　analysis.　The　calculation　accuracy　of　the　series　of　prediction　equations　proposed　in　this　paper　were　verified　by　the　collected　literature　data.　Meanwhile,　comparing　the　prediction　equations　proposed　in　this　paper　with　the　existing　equations　in　the　literature,　the　equations　proposed　in　this　paper　were　better.　©　2023　Elsevier　Ltd