diff --git a/cgrcompute/components/courserecommendation.py b/cgrcompute/components/courserecommendation.py
index 31728f2..66094c0 100644
--- a/cgrcompute/components/courserecommendation.py
+++ b/cgrcompute/components/courserecommendation.py
@@ -1,5 +1,6 @@
 from scipy.sparse import lil_matrix
-from sklearn.metrics.pairwise import cosine_similarity
+from sklearn.metrics.pairwise import cosine_similarity,linear_kernel
+from sklearn.preprocessing import normalize
 from cgrcompute.components.external import ElasticService, get_mongo_service
 from typing import Hashable
 from logging import getLogger
@@ -30,6 +31,67 @@ def train(observations: list[set[Hashable]]) -> 'CosineSimRecommendationModel':
             neigh = sorted(neigh, key=lambda x: x[1])[-300:]
             ccmtx[cid] = dict(neigh)
         return CosineSimRecommendationModel(ccmtx)
+    @staticmethod
+    def train32(observations: list[set[Hashable]]) -> 'CosineSimRecommendationModel':
+        items = list(set(c for o in observations for c in o))
+        itemidx = dict((c, i) for (i, c) in enumerate(items))
+        itemobsv = lil_matrix((len(items), len(observations)),dtype='float32')
+        for (i, o) in enumerate(observations):
+            for c in o:
+                itemobsv[itemidx[c], i] = 1
+        sim = cosine_similarity(itemobsv.astype('float32'), dense_output=False)
+        ccmtx = dict()
+        for (i, cid) in enumerate(items):
+            neigh = []
+            for c in sim.getrow(i).nonzero()[1]:
+                neigh.append((items[c], sim[i, c]))
+            neigh = sorted(neigh, key=lambda x: x[1])[-300:]
+            ccmtx[cid] = dict(neigh)
+        return CosineSimRecommendationModel(ccmtx)
+    @staticmethod
+    def train16(observations: list[set[Hashable]]) -> 'CosineSimRecommendationModel':
+        items = list(set(c for o in observations for c in o))
+        itemidx = dict((c, i) for (i, c) in enumerate(items))
+        itemobsv = lil_matrix((len(items), len(observations)),dtype='float32')
+        for (i, o) in enumerate(observations):
+            for c in o:
+                itemobsv[itemidx[c], i] = 1
+        size = itemobsv.shape[0]
+        chunk_size = 500
+        ccmtx = dict()
+        for start in range(0,size,chunk_size):
+            end = min(start+chunk_size,size)
+            sim = cosine_similarity(itemobsv[start:end],itemobsv,dense_output=False).astype('float16')
+            for (i, cid) in enumerate(items[start:end]):
+                neigh = []
+                for c in sim.getrow(i).nonzero()[1]:
+                    neigh.append((items[c], sim[i, c]))
+                neigh = sorted(neigh, key=lambda x: x[1])[-300:]
+                ccmtx[cid] = dict(neigh)
+        return CosineSimRecommendationModel(ccmtx)
+    @staticmethod
+    def train8(observations: list[set[Hashable]]) -> 'CosineSimRecommendationModel':
+        items = list(set(c for o in observations for c in o))
+        itemidx = dict((c, i) for (i, c) in enumerate(items))
+        itemobsv = lil_matrix((len(items), len(observations)),dtype='uint8')
+        for (i, o) in enumerate(observations):
+            for c in o:
+                itemobsv[itemidx[c], i] = 1
+        size = itemobsv.shape[0]
+        chunk_size = 500
+        cnorm = normalize(itemobsv)
+        cnorm *=16
+        ccmtx = dict()
+        for start in range(0,size,chunk_size):
+            end = min(start+chunk_size,size)
+            sim = linear_kernel(cnorm[start:end],cnorm,dense_output=False).astype('uint8')
+            for (i, cid) in enumerate(items[start:end]):
+                neigh = []
+                for c in sim.getrow(i).nonzero()[1]:
+                    neigh.append((items[c], sim[i, c]))
+                neigh = sorted(neigh, key=lambda x: x[1])[-300:]
+                ccmtx[cid] = dict(neigh)
+        return CosineSimRecommendationModel(ccmtx)
 
     def infer(self, selected_item: list[Hashable]) -> dict[Hashable, float]:
         d = dict()
@@ -48,6 +110,9 @@ class CourseRecommendationModel:
     
     def __init__(self):
         self.model = None
+        self.model32 = None
+        self.model16 = None
+        self.model8 = None
         self.logger = getLogger('CourseRecommendationModel')
 
     def populate(self):
@@ -55,10 +120,20 @@ def populate(self):
         obsv = self.downloadobsvdata()
         self.logger.info("Download completed {}. Start training".format(time.time()))
         self.model = CosineSimRecommendationModel.train(obsv)
+        self.model32 = CosineSimRecommendationModel.train32(obsv)
+        self.model16 = CosineSimRecommendationModel.train16(obsv)
+        self.model8 = CosineSimRecommendationModel.train8(obsv)
         self.logger.info("Training completed {}".format(time.time()))
 
-    def infer(self, selected_courses):
-        res = self.model.infer(selected_courses)
+    def infer(self, selected_courses,variant):
+        if variant =="COSINE":
+            res = self.model.infer(selected_courses)
+        elif variant =="COSINE32":
+            res = self.model32.infer(selected_courses)
+        elif variant =="COSINE16":
+            res = self.model16.infer(selected_courses)
+        elif variant =="COSINE8":
+            res = self.model8.infer(selected_courses)
         res = sorted(res.items(), key=lambda x:-x[1])
         return [course for course, score in res][:300]
     
@@ -99,8 +174,8 @@ def recommend_course(req: grpcmsg.CourseRecommendationRequest, cache: SharableCa
     res = []
     if req.variant == 'RANDOM':
         res = model.random_infer()
-    elif req.variant == 'COSINE':
-        res = model.infer([(e.semesterKey.studyProgram, e.courseNo) for e in req.selectedCourses])
+    elif req.variant in ['COSINE','COSINE32','COSINE16','COSINE8']:
+        res = model.infer([(e.semesterKey.studyProgram, e.courseNo) for e in req.selectedCourses],req.variant)
     else:
         raise Exception('{} variant is invalid'.format(req.variant))
     enriched_res = []