The　SVM　needs　use　approximation　accuracy　Ε　,　however　the　LSJSVM　(least　square　support　vector　machine)　doesn＇t　need　Ε　.　According　to　these　characteristics,　this　paper　studies　the　fitting　and　generalization　capabilities　of　models　that　LS-SVM　and　SVM　establishes　for　the　penicillin　fermentation　processes　respectively.　An　improved　GA　selects　the　parameter　values　for　LS-SVM　and　SVM　respectively.　The　experiment　shows　the　model　based　on　LSJSVM　possesses　the　strong　capabilities　of　fitting　and　generalization.　If　Ε　is　too　large,　the　capabilities　of　fitting　and　generalization　of　model　based　on　SVM　are　not　high;　if　Ε　is　too　small,　the　capabilities　of　fitting　and　generalization　are　relatively　high,　but　the　modeling　process　demands　more　long　time.　So,　the　LS-SVM　is　more　suitable　for　modeling　in　fermentation　processes.　©　2009　IEEE.