diff --git a/spacy/_ml.py b/spacy/_ml.py
index d3bb903e7..f9894cd54 100644
--- a/spacy/_ml.py
+++ b/spacy/_ml.py
@@ -93,7 +93,10 @@ class PrecomputableMaxouts(Model):
         # W: (f, o, p, i)
         # b: (o, p)
 
-        Yfp = numpy.einsum('bi,fopi->fbop', X, self.W)
+        # Yfp = numpy.einsum('bi,fopi->fbop', X, self.W)
+        Yfp = self.ops.xp.tensordot(X, self.W,
+                axes=[[1], [3]]).transpose((1, 0, 2, 3))
+        Yfp = self.ops.xp.ascontiguousarray(Yfp)
         Yfp += self.b
         Yf = self.ops.allocate((self.nF, X.shape[0], self.nO))
         which = self.ops.allocate((self.nF, X.shape[0], self.nO), dtype='i')
@@ -106,8 +109,11 @@ class PrecomputableMaxouts(Model):
             for i in range(self.nF):
                 dYp += self.ops.backprop_maxout(dY, which[i], self.nP)
 
-            dXf = numpy.einsum('bop,fopi->bfi', dYp, self.W)
-            dW = numpy.einsum('bop,bfi->fopi', dYp, Xf)
+            #dXf = numpy.einsum('bop,fopi->bfi', dYp, self.W)
+            dXf = self.ops.xp.tensordot(dYp, self.W, axes=[[1,2], [1,2]])
+            #dW = numpy.einsum('bfi,bop->fopi', Xf, dYp)
+            dW = self.ops.xp.tensordot(Xf, dYp, axes=[[0], [0]])
+            dW = dW.transpose((0, 2, 3, 1))
             db = dYp.sum(axis=0)
 
             self.d_W += dW
diff --git a/spacy/pipeline.pyx b/spacy/pipeline.pyx
index 61c71c2bb..c357397f8 100644
--- a/spacy/pipeline.pyx
+++ b/spacy/pipeline.pyx
@@ -21,7 +21,7 @@ class TokenVectorEncoder(object):
     '''Assign position-sensitive vectors to tokens, using a CNN or RNN.'''
     def __init__(self, vocab, **cfg):
         self.vocab = vocab
-        self.model = build_tok2vec(vocab.lang, 64, **cfg)
+        self.model = build_tok2vec(vocab.lang, **cfg)
         self.tagger = chain(
                         self.model,
                         Softmax(self.vocab.morphology.n_tags))
diff --git a/spacy/syntax/parser.pyx b/spacy/syntax/parser.pyx
index 4989a1fb3..76f16f881 100644
--- a/spacy/syntax/parser.pyx
+++ b/spacy/syntax/parser.pyx
@@ -89,7 +89,7 @@ def get_greedy_model_for_batch(tokvecs, TransitionSystem moves, upper_model, low
         for i, offset in enumerate(offsets):
             adjusted_ids[i] *= token_ids[i] >= 0
             adjusted_ids[i] += offset
-        features = upper_model.ops.allocate((len(states), 64), dtype='f')
+        features = upper_model.ops.allocate((len(states), lower_model.nO), dtype='f')
         for i in range(len(states)):
             for j, tok_i in enumerate(adjusted_ids[i]):
                 if tok_i >= 0:
@@ -222,7 +222,7 @@ cdef class Parser:
         nr_context_tokens = StateClass.nr_context_tokens(nF, nB, nS, nL, nR)
 
         upper = chain(Maxout(width, width), Maxout(self.moves.n_moves, width))
-        lower = PrecomputableMaxouts(width, nF=nr_context_tokens, nI=width)
+        lower = PrecomputableMaxouts(width, nF=nr_context_tokens, nI=width*2)
         return upper, lower
 
     def __call__(self, Doc tokens):