diff --git a/setup.py b/setup.py
index 0a3384ed5..02d4fe0d9 100755
--- a/setup.py
+++ b/setup.py
@@ -36,6 +36,7 @@ MOD_NAMES = [
     'spacy.syntax.transition_system',
     'spacy.syntax.arc_eager',
     'spacy.syntax._parse_features',
+    'spacy.syntax._beam_utils',
     'spacy.gold',
     'spacy.tokens.doc',
     'spacy.tokens.span',
diff --git a/spacy/_ml.py b/spacy/_ml.py
index f1ded666e..91b530fad 100644
--- a/spacy/_ml.py
+++ b/spacy/_ml.py
@@ -5,10 +5,12 @@ from thinc.neural._classes.hash_embed import HashEmbed
 from thinc.neural.ops import NumpyOps, CupyOps
 from thinc.neural.util import get_array_module
 import random
+import cytoolz
 
 from thinc.neural._classes.convolution import ExtractWindow
 from thinc.neural._classes.static_vectors import StaticVectors
 from thinc.neural._classes.batchnorm import BatchNorm
+from thinc.neural._classes.layernorm import LayerNorm as LN
 from thinc.neural._classes.resnet import Residual
 from thinc.neural import ReLu
 from thinc.neural._classes.selu import SELU
@@ -19,10 +21,12 @@ from thinc.api import FeatureExtracter, with_getitem
 from thinc.neural.pooling import Pooling, max_pool, mean_pool, sum_pool
 from thinc.neural._classes.attention import ParametricAttention
 from thinc.linear.linear import LinearModel
-from thinc.api import uniqued, wrap
+from thinc.api import uniqued, wrap, flatten_add_lengths
+
 
 from .attrs import ID, ORTH, LOWER, NORM, PREFIX, SUFFIX, SHAPE, TAG, DEP
 from .tokens.doc import Doc
+from . import util
 
 import numpy
 import io
@@ -53,6 +57,27 @@ def _logistic(X, drop=0.):
     return Y, logistic_bwd
 
 
+@layerize
+def add_tuples(X, drop=0.):
+    """Give inputs of sequence pairs, where each sequence is (vals, length),
+    sum the values, returning a single sequence.
+
+    If input is:
+    ((vals1, length), (vals2, length)
+    Output is:
+    (vals1+vals2, length)
+
+    vals are a single tensor for the whole batch.
+    """
+    (vals1, length1), (vals2, length2) = X
+    assert length1 == length2
+
+    def add_tuples_bwd(dY, sgd=None):
+        return (dY, dY)
+
+    return (vals1+vals2, length), add_tuples_bwd
+
+
 def _zero_init(model):
     def _zero_init_impl(self, X, y):
         self.W.fill(0)
@@ -61,6 +86,7 @@ def _zero_init(model):
         model.W.fill(0.)
     return model
 
+
 @layerize
 def _preprocess_doc(docs, drop=0.):
     keys = [doc.to_array([LOWER]) for doc in docs]
@@ -72,7 +98,6 @@ def _preprocess_doc(docs, drop=0.):
     return (keys, vals, lengths), None
 
 
-
 def _init_for_precomputed(W, ops):
     if (W**2).sum() != 0.:
         return
@@ -80,6 +105,7 @@ def _init_for_precomputed(W, ops):
     ops.xavier_uniform_init(reshaped)
     W[:] = reshaped.reshape(W.shape)
 
+
 @describe.on_data(_set_dimensions_if_needed)
 @describe.attributes(
     nI=Dimension("Input size"),
@@ -184,25 +210,36 @@ class PrecomputableMaxouts(Model):
         return Yfp, backward
 
 
+def drop_layer(layer, factor=2.):
+    def drop_layer_fwd(X, drop=0.):
+        drop *= factor
+        mask = layer.ops.get_dropout_mask((1,), drop)
+        if mask is None or mask > 0:
+            return layer.begin_update(X, drop=drop)
+        else:
+            return X, lambda dX, sgd=None: dX
+    return wrap(drop_layer_fwd, layer)
+
+
 def Tok2Vec(width, embed_size, preprocess=None):
-    cols = [ID, NORM, PREFIX, SUFFIX, SHAPE]
+    cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
     with Model.define_operators({'>>': chain, '|': concatenate, '**': clone, '+': add}):
-        norm = get_col(cols.index(NORM))   >> HashEmbed(width, embed_size, name='embed_lower')
+        norm = get_col(cols.index(NORM))     >> HashEmbed(width, embed_size, name='embed_lower')
         prefix = get_col(cols.index(PREFIX)) >> HashEmbed(width, embed_size//2, name='embed_prefix')
         suffix = get_col(cols.index(SUFFIX)) >> HashEmbed(width, embed_size//2, name='embed_suffix')
         shape = get_col(cols.index(SHAPE))   >> HashEmbed(width, embed_size//2, name='embed_shape')
 
-        embed = (norm | prefix | suffix | shape )
+        embed = (norm | prefix | suffix | shape ) >> Maxout(width, width*4, pieces=3)
         tok2vec = (
             with_flatten(
                 asarray(Model.ops, dtype='uint64')
-                >> embed
-                >> Maxout(width, width*4, pieces=3)
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3))
-                >> Residual(ExtractWindow(nW=1) >> Maxout(width, width*3)),
-            pad=4)
+                >> uniqued(embed, column=5)
+                >> drop_layer(
+                    Residual(
+                        (ExtractWindow(nW=1) >> ReLu(width, width*3))
+                    )
+                ) ** 4, pad=4
+            )
         )
         if preprocess not in (False, None):
             tok2vec = preprocess >> tok2vec
@@ -297,7 +334,8 @@ def zero_init(model):
 
 
 def doc2feats(cols=None):
-    cols = [ID, NORM, PREFIX, SUFFIX, SHAPE]
+    if cols is None:
+        cols = [ID, NORM, PREFIX, SUFFIX, SHAPE, ORTH]
     def forward(docs, drop=0.):
         feats = []
         for doc in docs:
@@ -323,6 +361,37 @@ def get_token_vectors(tokens_attrs_vectors, drop=0.):
     return vectors, backward
 
 
+def fine_tune(embedding, combine=None):
+    if combine is not None:
+        raise NotImplementedError(
+            "fine_tune currently only supports addition. Set combine=None")
+    def fine_tune_fwd(docs_tokvecs, drop=0.):
+        docs, tokvecs = docs_tokvecs
+        lengths = model.ops.asarray([len(doc) for doc in docs], dtype='i')
+
+        vecs, bp_vecs = embedding.begin_update(docs, drop=drop)
+        flat_tokvecs = embedding.ops.flatten(tokvecs)
+        flat_vecs = embedding.ops.flatten(vecs)
+        output = embedding.ops.unflatten(
+                   (model.mix[0] * flat_vecs + model.mix[1] * flat_tokvecs),
+                    lengths)
+
+        def fine_tune_bwd(d_output, sgd=None):
+            bp_vecs(d_output, sgd=sgd)
+            flat_grad = model.ops.flatten(d_output)
+            model.d_mix[1] += flat_tokvecs.dot(flat_grad.T).sum()
+            model.d_mix[0] += flat_vecs.dot(flat_grad.T).sum()
+            if sgd is not None:
+                sgd(model._mem.weights, model._mem.gradient, key=model.id)
+            return d_output
+        return output, fine_tune_bwd
+    model = wrap(fine_tune_fwd, embedding)
+    model.mix = model._mem.add((model.id, 'mix'), (2,))
+    model.mix.fill(1.)
+    model.d_mix = model._mem.add_gradient((model.id, 'd_mix'), (model.id, 'mix'))
+    return model
+
+
 @layerize
 def flatten(seqs, drop=0.):
     if isinstance(seqs[0], numpy.ndarray):
@@ -369,6 +438,27 @@ def preprocess_doc(docs, drop=0.):
     vals = ops.allocate(keys.shape[0]) + 1
     return (keys, vals, lengths), None
 
+def getitem(i):
+    def getitem_fwd(X, drop=0.):
+        return X[i], None
+    return layerize(getitem_fwd)
+
+def build_tagger_model(nr_class, token_vector_width, **cfg):
+    embed_size = util.env_opt('embed_size', 7500)
+    with Model.define_operators({'>>': chain, '+': add}):
+        # Input: (doc, tensor) tuples
+        private_tok2vec = Tok2Vec(token_vector_width, embed_size, preprocess=doc2feats())
+
+        model = (
+            fine_tune(private_tok2vec)
+            >> with_flatten(
+                Maxout(token_vector_width, token_vector_width)
+                >> Softmax(nr_class, token_vector_width)
+            )
+        )
+    model.nI = None
+    return model
+
 
 def build_text_classifier(nr_class, width=64, **cfg):
     nr_vector = cfg.get('nr_vector', 200)
@@ -383,7 +473,7 @@ def build_text_classifier(nr_class, width=64, **cfg):
             >> _flatten_add_lengths
             >> with_getitem(0,
                 uniqued(
-                  (embed_lower | embed_prefix | embed_suffix | embed_shape) 
+                  (embed_lower | embed_prefix | embed_suffix | embed_shape)
                   >> Maxout(width, width+(width//2)*3))
                 >> Residual(ExtractWindow(nW=1) >> ReLu(width, width*3))
                 >> Residual(ExtractWindow(nW=1) >> ReLu(width, width*3))
@@ -404,7 +494,7 @@ def build_text_classifier(nr_class, width=64, **cfg):
             >> zero_init(Affine(nr_class, nr_class*2, drop_factor=0.0))
             >> logistic
         )
- 
+
     model.lsuv = False
     return model
 
diff --git a/spacy/cli/convert.py b/spacy/cli/convert.py
index a0a76e5ec..fef6753e6 100644
--- a/spacy/cli/convert.py
+++ b/spacy/cli/convert.py
@@ -21,10 +21,10 @@ CONVERTERS = {
 @plac.annotations(
     input_file=("input file", "positional", None, str),
     output_dir=("output directory for converted file", "positional", None, str),
-    n_sents=("Number of sentences per doc", "option", "n", float),
+    n_sents=("Number of sentences per doc", "option", "n", int),
     morphology=("Enable appending morphology to tags", "flag", "m", bool)
 )
-def convert(cmd, input_file, output_dir, n_sents, morphology):
+def convert(cmd, input_file, output_dir, n_sents=1, morphology=False):
     """
     Convert files into JSON format for use with train command and other
     experiment management functions.
diff --git a/spacy/cli/train.py b/spacy/cli/train.py
index af028dae5..04aac8319 100644
--- a/spacy/cli/train.py
+++ b/spacy/cli/train.py
@@ -91,15 +91,14 @@ def train(cmd, lang, output_dir, train_data, dev_data, n_iter=20, n_sents=0,
                 for batch in minibatch(train_docs, size=batch_sizes):
                     docs, golds = zip(*batch)
                     nlp.update(docs, golds, sgd=optimizer,
-                               drop=next(dropout_rates), losses=losses)
+                               drop=next(dropout_rates), losses=losses,
+                               update_tensors=True)
                     pbar.update(sum(len(doc) for doc in docs))
 
             with nlp.use_params(optimizer.averages):
                 util.set_env_log(False)
                 epoch_model_path = output_path / ('model%d' % i)
                 nlp.to_disk(epoch_model_path)
-                with (output_path / ('model%d.pickle' % i)).open('wb') as file_:
-                    dill.dump(nlp, file_, -1)
                 nlp_loaded = lang_class(pipeline=pipeline)
                 nlp_loaded = nlp_loaded.from_disk(epoch_model_path)
                 scorer = nlp_loaded.evaluate(
diff --git a/spacy/language.py b/spacy/language.py
index 0284c4636..cb679a2bc 100644
--- a/spacy/language.py
+++ b/spacy/language.py
@@ -277,7 +277,8 @@ class Language(object):
     def make_doc(self, text):
         return self.tokenizer(text)
 
-    def update(self, docs, golds, drop=0., sgd=None, losses=None):
+    def update(self, docs, golds, drop=0., sgd=None, losses=None,
+            update_tensors=False):
         """Update the models in the pipeline.
 
         docs (iterable): A batch of `Doc` objects.
@@ -304,14 +305,17 @@ class Language(object):
             grads[key] = (W, dW)
         pipes = list(self.pipeline[1:])
         random.shuffle(pipes)
+        tokvecses, bp_tokvecses = tok2vec.model.begin_update(feats, drop=drop)
+        all_d_tokvecses = [tok2vec.model.ops.allocate(tv.shape) for tv in tokvecses]
         for proc in pipes:
             if not hasattr(proc, 'update'):
                 continue
-            tokvecses, bp_tokvecses = tok2vec.model.begin_update(feats, drop=drop)
             d_tokvecses = proc.update((docs, tokvecses), golds,
                                       drop=drop, sgd=get_grads, losses=losses)
-            if d_tokvecses is not None:
-                bp_tokvecses(d_tokvecses, sgd=sgd)
+            if update_tensors and d_tokvecses is not None:
+                for i, d_tv in enumerate(d_tokvecses):
+                    all_d_tokvecses[i] += d_tv
+        bp_tokvecses(all_d_tokvecses, sgd=sgd)
         for key, (W, dW) in grads.items():
             sgd(W, dW, key=key)
         # Clear the tensor variable, to free GPU memory.
@@ -381,9 +385,18 @@ class Language(object):
         return optimizer
 
     def evaluate(self, docs_golds):
-        docs, golds = zip(*docs_golds)
         scorer = Scorer()
-        for doc, gold in zip(self.pipe(docs, batch_size=32), golds):
+        docs, golds = zip(*docs_golds)
+        docs = list(docs)
+        golds = list(golds)
+        for pipe in self.pipeline:
+            if not hasattr(pipe, 'pipe'):
+                for doc in docs:
+                    pipe(doc)
+            else:
+                docs = list(pipe.pipe(docs))
+        assert len(docs) == len(golds)
+        for doc, gold in zip(docs, golds):
             scorer.score(doc, gold)
             doc.tensor = None
         return scorer
diff --git a/spacy/pipeline.pyx b/spacy/pipeline.pyx
index 947f0a1f1..634d3e4b5 100644
--- a/spacy/pipeline.pyx
+++ b/spacy/pipeline.pyx
@@ -42,7 +42,7 @@ from .compat import json_dumps
 
 from .attrs import ID, LOWER, PREFIX, SUFFIX, SHAPE, TAG, DEP, POS
 from ._ml import rebatch, Tok2Vec, flatten, get_col, doc2feats
-from ._ml import build_text_classifier
+from ._ml import build_text_classifier, build_tagger_model
 from .parts_of_speech import X
 
 
@@ -138,7 +138,7 @@ class TokenVectorEncoder(BaseThincComponent):
     name = 'tensorizer'
 
     @classmethod
-    def Model(cls, width=128, embed_size=7500, **cfg):
+    def Model(cls, width=128, embed_size=4000, **cfg):
         """Create a new statistical model for the class.
 
         width (int): Output size of the model.
@@ -253,23 +253,25 @@ class NeuralTagger(BaseThincComponent):
         self.cfg = dict(cfg)
 
     def __call__(self, doc):
-        tags = self.predict([doc.tensor])
+        tags = self.predict(([doc], [doc.tensor]))
         self.set_annotations([doc], tags)
         return doc
 
     def pipe(self, stream, batch_size=128, n_threads=-1):
         for docs in cytoolz.partition_all(batch_size, stream):
+            docs = list(docs)
             tokvecs = [d.tensor for d in docs]
-            tag_ids = self.predict(tokvecs)
+            tag_ids = self.predict((docs, tokvecs))
             self.set_annotations(docs, tag_ids)
             yield from docs
 
-    def predict(self, tokvecs):
-        scores = self.model(tokvecs)
+    def predict(self, docs_tokvecs):
+        scores = self.model(docs_tokvecs)
         scores = self.model.ops.flatten(scores)
         guesses = scores.argmax(axis=1)
         if not isinstance(guesses, numpy.ndarray):
             guesses = guesses.get()
+        tokvecs = docs_tokvecs[1]
         guesses = self.model.ops.unflatten(guesses,
                     [tv.shape[0] for tv in tokvecs])
         return guesses
@@ -282,6 +284,8 @@ class NeuralTagger(BaseThincComponent):
         cdef Vocab vocab = self.vocab
         for i, doc in enumerate(docs):
             doc_tag_ids = batch_tag_ids[i]
+            if hasattr(doc_tag_ids, 'get'):
+                doc_tag_ids = doc_tag_ids.get()
             for j, tag_id in enumerate(doc_tag_ids):
                 # Don't clobber preset POS tags
                 if doc.c[j].tag == 0 and doc.c[j].pos == 0:
@@ -294,8 +298,7 @@ class NeuralTagger(BaseThincComponent):
 
         if self.model.nI is None:
             self.model.nI = tokvecs[0].shape[1]
-
-        tag_scores, bp_tag_scores = self.model.begin_update(tokvecs, drop=drop)
+        tag_scores, bp_tag_scores = self.model.begin_update(docs_tokvecs, drop=drop)
         loss, d_tag_scores = self.get_loss(docs, golds, tag_scores)
 
         d_tokvecs = bp_tag_scores(d_tag_scores, sgd=sgd)
@@ -346,10 +349,8 @@ class NeuralTagger(BaseThincComponent):
 
     @classmethod
     def Model(cls, n_tags, token_vector_width):
-        return with_flatten(
-            chain(Maxout(token_vector_width, token_vector_width),
-                  Softmax(n_tags, token_vector_width)))
-
+        return build_tagger_model(n_tags, token_vector_width)
+ 
     def use_params(self, params):
         with self.model.use_params(params):
             yield
@@ -432,7 +433,7 @@ class NeuralLabeller(NeuralTagger):
 
     @property
     def labels(self):
-        return self.cfg.get('labels', {})
+        return self.cfg.setdefault('labels', {})
 
     @labels.setter
     def labels(self, value):
@@ -455,10 +456,8 @@ class NeuralLabeller(NeuralTagger):
 
     @classmethod
     def Model(cls, n_tags, token_vector_width):
-        return with_flatten(
-            chain(Maxout(token_vector_width, token_vector_width),
-                  Softmax(n_tags, token_vector_width)))
-
+        return build_tagger_model(n_tags, token_vector_width)
+    
     def get_loss(self, docs, golds, scores):
         scores = self.model.ops.flatten(scores)
         cdef int idx = 0
diff --git a/spacy/syntax/_beam_utils.pyx b/spacy/syntax/_beam_utils.pyx
new file mode 100644
index 000000000..e77036e55
--- /dev/null
+++ b/spacy/syntax/_beam_utils.pyx
@@ -0,0 +1,273 @@
+# cython: infer_types=True
+# cython: profile=True
+cimport numpy as np
+import numpy
+from cpython.ref cimport PyObject, Py_INCREF, Py_XDECREF
+from thinc.extra.search cimport Beam
+from thinc.extra.search import MaxViolation
+from thinc.typedefs cimport hash_t, class_t
+
+from .transition_system cimport TransitionSystem, Transition
+from .stateclass cimport StateClass
+from ..gold cimport GoldParse
+from ..tokens.doc cimport Doc
+
+
+# These are passed as callbacks to thinc.search.Beam
+cdef int _transition_state(void* _dest, void* _src, class_t clas, void* _moves) except -1:
+    dest = <StateClass>_dest
+    src = <StateClass>_src
+    moves = <const Transition*>_moves
+    dest.clone(src)
+    moves[clas].do(dest.c, moves[clas].label)
+
+
+cdef int _check_final_state(void* _state, void* extra_args) except -1:
+    return (<StateClass>_state).is_final()
+
+
+def _cleanup(Beam beam):
+    for i in range(beam.width):
+        Py_XDECREF(<PyObject*>beam._states[i].content)
+        Py_XDECREF(<PyObject*>beam._parents[i].content)
+
+
+cdef hash_t _hash_state(void* _state, void* _) except 0:
+    state = <StateClass>_state
+    if state.c.is_final():
+        return 1
+    else:
+        return state.c.hash()
+
+
+cdef class ParserBeam(object):
+    cdef public TransitionSystem moves
+    cdef public object states
+    cdef public object golds
+    cdef public object beams
+
+    def __init__(self, TransitionSystem moves, states, golds,
+            int width=4, float density=0.001):
+        self.moves = moves
+        self.states = states
+        self.golds = golds
+        self.beams = []
+        cdef Beam beam
+        cdef StateClass state, st
+        for state in states:
+            beam = Beam(self.moves.n_moves, width, density)
+            beam.initialize(self.moves.init_beam_state, state.c.length, state.c._sent)
+            for i in range(beam.width):
+                st = <StateClass>beam.at(i)
+                st.c.offset = state.c.offset
+            self.beams.append(beam)
+
+    def __dealloc__(self):
+        if self.beams is not None:
+            for beam in self.beams:
+                if beam is not None:
+                    _cleanup(beam)
+
+    @property
+    def is_done(self):
+        return all(b.is_done for b in self.beams)
+
+    def __getitem__(self, i):
+        return self.beams[i]
+
+    def __len__(self):
+        return len(self.beams)
+
+    def advance(self, scores, follow_gold=False):
+        cdef Beam beam
+        for i, beam in enumerate(self.beams):
+            if beam.is_done or not scores[i].size:
+                continue
+            self._set_scores(beam, scores[i])
+            if self.golds is not None:
+                self._set_costs(beam, self.golds[i], follow_gold=follow_gold)
+            if follow_gold:
+                assert self.golds is not None
+                beam.advance(_transition_state, NULL, <void*>self.moves.c)
+            else:
+                beam.advance(_transition_state, _hash_state, <void*>self.moves.c)
+            beam.check_done(_check_final_state, NULL)
+            if beam.is_done:
+                for j in range(beam.size):
+                    if is_gold(<StateClass>beam.at(j), self.golds[i], self.moves.strings):
+                        beam._states[j].loss = 0.0
+                    elif beam._states[j].loss == 0.0:
+                        beam._states[j].loss = 1.0
+
+    def _set_scores(self, Beam beam, float[:, ::1] scores):
+        cdef float* c_scores = &scores[0, 0]
+        for i in range(beam.size):
+            state = <StateClass>beam.at(i)
+            if not state.is_final():
+                for j in range(beam.nr_class):
+                    beam.scores[i][j] = c_scores[i * beam.nr_class + j]
+                self.moves.set_valid(beam.is_valid[i], state.c)
+
+    def _set_costs(self, Beam beam, GoldParse gold, int follow_gold=False):
+        for i in range(beam.size):
+            state = <StateClass>beam.at(i)
+            if not state.c.is_final():
+                self.moves.set_costs(beam.is_valid[i], beam.costs[i], state, gold)
+                if follow_gold:
+                    for j in range(beam.nr_class):
+                        if beam.costs[i][j] >= 1:
+                            beam.is_valid[i][j] = 0
+
+
+def is_gold(StateClass state, GoldParse gold, strings):
+    predicted = set()
+    truth = set()
+    for i in range(gold.length):
+        if gold.cand_to_gold[i] is None:
+            continue
+        if state.safe_get(i).dep:
+            predicted.add((i, state.H(i), strings[state.safe_get(i).dep]))
+        else:
+            predicted.add((i, state.H(i), 'ROOT'))
+        id_, word, tag, head, dep, ner = gold.orig_annot[gold.cand_to_gold[i]]
+        truth.add((id_, head, dep))
+    return truth == predicted
+
+
+def get_token_ids(states, int n_tokens):
+    cdef StateClass state
+    cdef np.ndarray ids = numpy.zeros((len(states), n_tokens),
+                                      dtype='int32', order='C')
+    c_ids = <int*>ids.data
+    for i, state in enumerate(states):
+        if not state.is_final():
+            state.c.set_context_tokens(c_ids, n_tokens)
+        else:
+            ids[i] = -1
+        c_ids += ids.shape[1]
+    return ids
+
+nr_update = 0
+def update_beam(TransitionSystem moves, int nr_feature, int max_steps,
+                states, tokvecs, golds,
+                state2vec, vec2scores, drop=0., sgd=None,
+                losses=None, int width=4, float density=0.001):
+    global nr_update
+    nr_update += 1
+    pbeam = ParserBeam(moves, states, golds,
+                       width=width, density=density)
+    gbeam = ParserBeam(moves, states, golds,
+                       width=width, density=0.0)
+    cdef StateClass state
+    beam_maps = []
+    backprops = []
+    violns = [MaxViolation() for _ in range(len(states))]
+    for t in range(max_steps):
+        # The beam maps let us find the right row in the flattened scores
+        # arrays for each state. States are identified by (example id, history).
+        # We keep a different beam map for each step (since we'll have a flat
+        # scores array for each step). The beam map will let us take the per-state
+        # losses, and compute the gradient for each (step, state, class).
+        beam_maps.append({})
+        # Gather all states from the two beams in a list. Some stats may occur
+        # in both beams. To figure out which beam each state belonged to,
+        # we keep two lists of indices, p_indices and g_indices
+        states, p_indices, g_indices = get_states(pbeam, gbeam, beam_maps[-1], nr_update)
+        if not states:
+            break
+        # Now that we have our flat list of states, feed them through the model
+        token_ids = get_token_ids(states, nr_feature)
+        vectors, bp_vectors = state2vec.begin_update(token_ids, drop=drop)
+        scores, bp_scores = vec2scores.begin_update(vectors, drop=drop)
+
+        # Store the callbacks for the backward pass
+        backprops.append((token_ids, bp_vectors, bp_scores))
+
+        # Unpack the flat scores into lists for the two beams. The indices arrays
+        # tell us which example and state the scores-row refers to.
+        p_scores = [numpy.ascontiguousarray(scores[indices], dtype='f') for indices in p_indices]
+        g_scores = [numpy.ascontiguousarray(scores[indices], dtype='f')  for indices in g_indices]
+        # Now advance the states in the beams. The gold beam is contrained to
+        # to follow only gold analyses.
+        pbeam.advance(p_scores)
+        gbeam.advance(g_scores, follow_gold=True)
+        # Track the "maximum violation", to use in the update.
+        for i, violn in enumerate(violns):
+            violn.check_crf(pbeam[i], gbeam[i])
+
+    # Only make updates if we have non-gold states
+    histories = [((v.p_hist + v.g_hist) if v.p_hist else []) for v in violns]
+    losses = [((v.p_probs + v.g_probs) if v.p_probs else []) for v in violns]
+    states_d_scores = get_gradient(moves.n_moves, beam_maps,
+                                   histories, losses)
+    assert len(states_d_scores) == len(backprops), (len(states_d_scores), len(backprops))
+    return states_d_scores, backprops
+
+
+def get_states(pbeams, gbeams, beam_map, nr_update):
+    seen = {}
+    states = []
+    p_indices = []
+    g_indices = []
+    cdef Beam pbeam, gbeam
+    assert len(pbeams) == len(gbeams)
+    for eg_id, (pbeam, gbeam) in enumerate(zip(pbeams, gbeams)):
+        p_indices.append([])
+        g_indices.append([])
+        if pbeam.loss > 0 and pbeam.min_score > gbeam.score:
+            continue
+        for i in range(pbeam.size):
+            state = <StateClass>pbeam.at(i)
+            if not state.is_final():
+                key = tuple([eg_id] + pbeam.histories[i])
+                seen[key] = len(states)
+                p_indices[-1].append(len(states))
+                states.append(state)
+        beam_map.update(seen)
+        for i in range(gbeam.size):
+            state = <StateClass>gbeam.at(i)
+            if not state.is_final():
+                key = tuple([eg_id] + gbeam.histories[i])
+                if key in seen:
+                    g_indices[-1].append(seen[key])
+                else:
+                    g_indices[-1].append(len(states))
+                    beam_map[key] = len(states)
+                    states.append(state)
+    p_idx = [numpy.asarray(idx, dtype='i') for idx in p_indices]
+    g_idx = [numpy.asarray(idx, dtype='i') for idx in g_indices]
+    return states, p_idx, g_idx
+
+
+def get_gradient(nr_class, beam_maps, histories, losses):
+    """
+    The global model assigns a loss to each parse. The beam scores
+    are additive, so the same gradient is applied to each action
+    in the history. This gives the gradient of a single *action*
+    for a beam state -- so we have "the gradient of loss for taking
+    action i given history H."
+
+    Histories: Each hitory is a list of actions
+    Each candidate has a history
+    Each beam has multiple candidates
+    Each batch has multiple beams
+    So history is list of lists of lists of ints
+    """
+    nr_step = len(beam_maps)
+    grads = []
+    for beam_map in beam_maps:
+        if beam_map:
+            grads.append(numpy.zeros((max(beam_map.values())+1, nr_class), dtype='f'))
+    assert len(histories) == len(losses)
+    for eg_id, hists in enumerate(histories):
+        for loss, hist in zip(losses[eg_id], hists):
+            key = tuple([eg_id])
+            for j, clas in enumerate(hist):
+                i = beam_maps[j][key]
+                # In step j, at state i action clas
+                # resulted in loss
+                grads[j][i, clas] += loss / len(histories)
+                key = key + tuple([clas])
+    return grads
+
+
diff --git a/spacy/syntax/_state.pxd b/spacy/syntax/_state.pxd
index c06851978..9aeeba441 100644
--- a/spacy/syntax/_state.pxd
+++ b/spacy/syntax/_state.pxd
@@ -37,6 +37,7 @@ cdef cppclass StateC:
         this.shifted = <bint*>calloc(length + (PADDING * 2), sizeof(bint))
         this._sent = <TokenC*>calloc(length + (PADDING * 2), sizeof(TokenC))
         this._ents = <Entity*>calloc(length + (PADDING * 2), sizeof(Entity))
+        this.offset = 0
         cdef int i
         for i in range(length + (PADDING * 2)):
             this._ents[i].end = -1
diff --git a/spacy/syntax/arc_eager.pyx b/spacy/syntax/arc_eager.pyx
index 29e8de0aa..9477449a5 100644
--- a/spacy/syntax/arc_eager.pyx
+++ b/spacy/syntax/arc_eager.pyx
@@ -385,6 +385,7 @@ cdef class ArcEager(TransitionSystem):
         for i in range(self.n_moves):
             if self.c[i].move == move and self.c[i].label == label:
                 return self.c[i]
+        return Transition(clas=0, move=MISSING, label=0)
 
     def move_name(self, int move, attr_t label):
         label_str = self.strings[label]
diff --git a/spacy/syntax/beam_parser.pyx b/spacy/syntax/beam_parser.pyx
index e96e28fcf..f4f66f9fb 100644
--- a/spacy/syntax/beam_parser.pyx
+++ b/spacy/syntax/beam_parser.pyx
@@ -34,6 +34,7 @@ from ._parse_features cimport CONTEXT_SIZE
 from ._parse_features cimport fill_context
 from .stateclass cimport StateClass
 from .parser cimport Parser
+from ._beam_utils import is_gold
 
 
 DEBUG = False
@@ -237,16 +238,3 @@ def _check_train_integrity(Beam pred, Beam gold, GoldParse gold_parse, Transitio
             raise Exception("Gold parse is not gold-standard")
 
 
-def is_gold(StateClass state, GoldParse gold, StringStore strings):
-    predicted = set()
-    truth = set()
-    for i in range(gold.length):
-        if gold.cand_to_gold[i] is None:
-            continue
-        if state.safe_get(i).dep:
-            predicted.add((i, state.H(i), strings[state.safe_get(i).dep]))
-        else:
-            predicted.add((i, state.H(i), 'ROOT'))
-        id_, word, tag, head, dep, ner = gold.orig_annot[gold.cand_to_gold[i]]
-        truth.add((id_, head, dep))
-    return truth == predicted
diff --git a/spacy/syntax/nn_parser.pxd b/spacy/syntax/nn_parser.pxd
index 524718965..7ff4b9f9f 100644
--- a/spacy/syntax/nn_parser.pxd
+++ b/spacy/syntax/nn_parser.pxd
@@ -14,8 +14,4 @@ cdef class Parser:
     cdef readonly TransitionSystem moves
     cdef readonly object cfg
 
-    cdef void _parse_step(self, StateC* state,
-            const float* feat_weights,
-            int nr_class, int nr_feat, int nr_piece) nogil
-
     #cdef int parseC(self, TokenC* tokens, int length, int nr_feat) nogil
diff --git a/spacy/syntax/nn_parser.pyx b/spacy/syntax/nn_parser.pyx
index 0b39e2216..a193c96a3 100644
--- a/spacy/syntax/nn_parser.pyx
+++ b/spacy/syntax/nn_parser.pyx
@@ -37,14 +37,17 @@ from preshed.maps cimport MapStruct
 from preshed.maps cimport map_get
 
 from thinc.api import layerize, chain, noop, clone
-from thinc.neural import Model, Affine, ELU, ReLu, Maxout
+from thinc.neural import Model, Affine, ReLu, Maxout
+from thinc.neural._classes.selu import SELU
+from thinc.neural._classes.layernorm import LayerNorm
 from thinc.neural.ops import NumpyOps, CupyOps
 from thinc.neural.util import get_array_module
 
 from .. import util
 from ..util import get_async, get_cuda_stream
 from .._ml import zero_init, PrecomputableAffine, PrecomputableMaxouts
-from .._ml import Tok2Vec, doc2feats, rebatch
+from .._ml import Tok2Vec, doc2feats, rebatch, fine_tune
+from .._ml import Residual, drop_layer
 from ..compat import json_dumps
 
 from . import _parse_features
@@ -59,8 +62,11 @@ from ..structs cimport TokenC
 from ..tokens.doc cimport Doc
 from ..strings cimport StringStore
 from ..gold cimport GoldParse
-from ..attrs cimport TAG, DEP
+from ..attrs cimport ID, TAG, DEP, ORTH, NORM, PREFIX, SUFFIX, TAG
+from . import _beam_utils
 
+USE_FINE_TUNE = True
+BEAM_PARSE = True
 
 def get_templates(*args, **kwargs):
     return []
@@ -232,11 +238,14 @@ cdef class Parser:
     Base class of the DependencyParser and EntityRecognizer.
     """
     @classmethod
-    def Model(cls, nr_class, token_vector_width=128, hidden_width=128, depth=1, **cfg):
+    def Model(cls, nr_class, token_vector_width=128, hidden_width=300, depth=1, **cfg):
         depth = util.env_opt('parser_hidden_depth', depth)
         token_vector_width = util.env_opt('token_vector_width', token_vector_width)
         hidden_width = util.env_opt('hidden_width', hidden_width)
         parser_maxout_pieces = util.env_opt('parser_maxout_pieces', 2)
+        embed_size = util.env_opt('embed_size', 4000)
+        tensors = fine_tune(Tok2Vec(token_vector_width, embed_size,
+                    preprocess=doc2feats()))
         if parser_maxout_pieces == 1:
             lower = PrecomputableAffine(hidden_width if depth >= 1 else nr_class,
                         nF=cls.nr_feature,
@@ -248,15 +257,10 @@ cdef class Parser:
                         nI=token_vector_width)
 
         with Model.use_device('cpu'):
-            if depth == 0:
-                upper = chain()
-                upper.is_noop = True
-            else:
-                upper = chain(
-                    clone(Maxout(hidden_width), (depth-1)),
-                    zero_init(Affine(nr_class, drop_factor=0.0))
-                )
-                upper.is_noop = False
+            upper = chain(
+                clone(Residual(ReLu(hidden_width)), (depth-1)),
+                zero_init(Affine(nr_class, drop_factor=0.0))
+            )
         # TODO: This is an unfortunate hack atm!
         # Used to set input dimensions in network.
         lower.begin_training(lower.ops.allocate((500, token_vector_width)))
@@ -268,7 +272,7 @@ cdef class Parser:
             'hidden_width': hidden_width,
             'maxout_pieces': parser_maxout_pieces
         }
-        return (lower, upper), cfg
+        return (tensors, lower, upper), cfg
 
     def __init__(self, Vocab vocab, moves=True, model=True, **cfg):
         """
@@ -344,17 +348,21 @@ cdef class Parser:
                 The number of threads with which to work on the buffer in parallel.
         Yields (Doc): Documents, in order.
         """
-        cdef StateClass parse_state
+        if BEAM_PARSE:
+            beam_width = 8
         cdef Doc doc
-        queue = []
+        cdef Beam beam
         for docs in cytoolz.partition_all(batch_size, docs):
             docs = list(docs)
-            tokvecs = [d.tensor for d in docs]
+            tokvecs = [doc.tensor for doc in docs]
             if beam_width == 1:
                 parse_states = self.parse_batch(docs, tokvecs)
             else:
-                parse_states = self.beam_parse(docs, tokvecs,
-                                    beam_width=beam_width, beam_density=beam_density)
+                beams = self.beam_parse(docs, tokvecs,
+                            beam_width=beam_width, beam_density=beam_density)
+                parse_states = []
+                for beam in beams:
+                    parse_states.append(<StateClass>beam.at(0))
             self.set_annotations(docs, parse_states)
             yield from docs
 
@@ -369,8 +377,12 @@ cdef class Parser:
             int nr_class, nr_feat, nr_piece, nr_dim, nr_state
         if isinstance(docs, Doc):
             docs = [docs]
+        if isinstance(tokvecses, np.ndarray):
+            tokvecses = [tokvecses]
 
         tokvecs = self.model[0].ops.flatten(tokvecses)
+        if USE_FINE_TUNE:
+            tokvecs += self.model[0].ops.flatten(self.model[0]((docs, tokvecses)))
 
         nr_state = len(docs)
         nr_class = self.moves.n_moves
@@ -394,27 +406,20 @@ cdef class Parser:
         cdef np.ndarray scores
         c_token_ids = <int*>token_ids.data
         c_is_valid = <int*>is_valid.data
-        cdef int has_hidden = not getattr(vec2scores, 'is_noop', False)
         while not next_step.empty():
-            if not has_hidden:
-                for i in cython.parallel.prange(
-                        next_step.size(), num_threads=6, nogil=True):
-                    self._parse_step(next_step[i],
-                        feat_weights, nr_class, nr_feat, nr_piece)
-            else:
-                for i in range(next_step.size()):
-                    st = next_step[i]
-                    st.set_context_tokens(&c_token_ids[i*nr_feat], nr_feat)
-                    self.moves.set_valid(&c_is_valid[i*nr_class], st)
+            for i in range(next_step.size()):
+                st = next_step[i]
+                st.set_context_tokens(&c_token_ids[i*nr_feat], nr_feat)
+                self.moves.set_valid(&c_is_valid[i*nr_class], st)
                 vectors = state2vec(token_ids[:next_step.size()])
-                scores = vec2scores(vectors)
-                c_scores = <float*>scores.data
-                for i in range(next_step.size()):
-                    st = next_step[i]
-                    guess = arg_max_if_valid(
-                        &c_scores[i*nr_class], &c_is_valid[i*nr_class], nr_class)
-                    action = self.moves.c[guess]
-                    action.do(st, action.label)
+            scores = vec2scores(vectors)
+            c_scores = <float*>scores.data
+            for i in range(next_step.size()):
+                st = next_step[i]
+                guess = arg_max_if_valid(
+                    &c_scores[i*nr_class], &c_is_valid[i*nr_class], nr_class)
+                action = self.moves.c[guess]
+                action.do(st, action.label)
             this_step, next_step = next_step, this_step
             next_step.clear()
             for st in this_step:
@@ -429,11 +434,15 @@ cdef class Parser:
         cdef int nr_class = self.moves.n_moves
         cdef StateClass stcls, output
         tokvecs = self.model[0].ops.flatten(tokvecses)
+        if USE_FINE_TUNE:
+            tokvecs += self.model[0].ops.flatten(self.model[0]((docs, tokvecses)))
         cuda_stream = get_cuda_stream()
         state2vec, vec2scores = self.get_batch_model(len(docs), tokvecs,
                                                      cuda_stream, 0.0)
         beams = []
         cdef int offset = 0
+        cdef int j = 0
+        cdef int k
         for doc in docs:
             beam = Beam(nr_class, beam_width, min_density=beam_density)
             beam.initialize(self.moves.init_beam_state, doc.length, doc.c)
@@ -446,44 +455,31 @@ cdef class Parser:
                 states = []
                 for i in range(beam.size):
                     stcls = <StateClass>beam.at(i)
-                    states.append(stcls)
+                    # This way we avoid having to score finalized states
+                    # We do have to take care to keep indexes aligned, though
+                    if not stcls.is_final():
+                        states.append(stcls)
                 token_ids = self.get_token_ids(states)
                 vectors = state2vec(token_ids)
                 scores = vec2scores(vectors)
+                j = 0
+                c_scores = <float*>scores.data
                 for i in range(beam.size):
                     stcls = <StateClass>beam.at(i)
                     if not stcls.is_final():
                         self.moves.set_valid(beam.is_valid[i], stcls.c)
-                        for j in range(nr_class):
-                            beam.scores[i][j] = scores[i, j]
+                        for k in range(nr_class):
+                            beam.scores[i][k] = c_scores[j * scores.shape[1] + k]
+                        j += 1
                 beam.advance(_transition_state, _hash_state, <void*>self.moves.c)
                 beam.check_done(_check_final_state, NULL)
             beams.append(beam)
         return beams
 
-    cdef void _parse_step(self, StateC* state,
-            const float* feat_weights,
-            int nr_class, int nr_feat, int nr_piece) nogil:
-        '''This only works with no hidden layers -- fast but inaccurate'''
-        #for i in cython.parallel.prange(next_step.size(), num_threads=4, nogil=True):
-        #    self._parse_step(next_step[i], feat_weights, nr_class, nr_feat)
-        token_ids = <int*>calloc(nr_feat, sizeof(int))
-        scores = <float*>calloc(nr_class * nr_piece, sizeof(float))
-        is_valid = <int*>calloc(nr_class, sizeof(int))
-
-        state.set_context_tokens(token_ids, nr_feat)
-        sum_state_features(scores,
-            feat_weights, token_ids, 1, nr_feat, nr_class * nr_piece)
-        self.moves.set_valid(is_valid, state)
-        guess = arg_maxout_if_valid(scores, is_valid, nr_class, nr_piece)
-        action = self.moves.c[guess]
-        action.do(state, action.label)
-
-        free(is_valid)
-        free(scores)
-        free(token_ids)
-
     def update(self, docs_tokvecs, golds, drop=0., sgd=None, losses=None):
+        if BEAM_PARSE:
+            return self.update_beam(docs_tokvecs, golds, drop=drop, sgd=sgd,
+                                    losses=losses)
         if losses is not None and self.name not in losses:
             losses[self.name] = 0.
         docs, tokvec_lists = docs_tokvecs
@@ -491,6 +487,10 @@ cdef class Parser:
         if isinstance(docs, Doc) and isinstance(golds, GoldParse):
             docs = [docs]
             golds = [golds]
+        if USE_FINE_TUNE:
+            my_tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop)
+            my_tokvecs = self.model[0].ops.flatten(my_tokvecs)
+            tokvecs += my_tokvecs
 
         cuda_stream = get_cuda_stream()
 
@@ -517,13 +517,13 @@ cdef class Parser:
             scores, bp_scores = vec2scores.begin_update(vector, drop=drop)
 
             d_scores = self.get_batch_loss(states, golds, scores)
-            d_vector = bp_scores(d_scores / d_scores.shape[0], sgd=sgd)
+            d_vector = bp_scores(d_scores, sgd=sgd)
             if drop != 0:
                 d_vector *= mask
 
             if isinstance(self.model[0].ops, CupyOps) \
             and not isinstance(token_ids, state2vec.ops.xp.ndarray):
-                # Move token_ids and d_vector to CPU, asynchronously
+                # Move token_ids and d_vector to GPU, asynchronously
                 backprops.append((
                     get_async(cuda_stream, token_ids),
                     get_async(cuda_stream, d_vector),
@@ -540,7 +540,55 @@ cdef class Parser:
                 break
         self._make_updates(d_tokvecs,
             backprops, sgd, cuda_stream)
-        return self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
+        d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, [len(d) for d in docs])
+        if USE_FINE_TUNE:
+            bp_my_tokvecs(d_tokvecs, sgd=sgd)
+        return d_tokvecs
+
+    def update_beam(self, docs_tokvecs, golds, drop=0., sgd=None, losses=None):
+        if losses is not None and self.name not in losses:
+            losses[self.name] = 0.
+        docs, tokvecs = docs_tokvecs
+        lengths = [len(d) for d in docs]
+        assert min(lengths) >= 1
+        tokvecs = self.model[0].ops.flatten(tokvecs)
+        if USE_FINE_TUNE:
+            my_tokvecs, bp_my_tokvecs = self.model[0].begin_update(docs_tokvecs, drop=drop)
+            my_tokvecs = self.model[0].ops.flatten(my_tokvecs)
+            tokvecs += my_tokvecs
+
+        states = self.moves.init_batch(docs)
+        for gold in golds:
+            self.moves.preprocess_gold(gold)
+
+        cuda_stream = get_cuda_stream()
+        state2vec, vec2scores = self.get_batch_model(len(states), tokvecs, cuda_stream, 0.0)
+
+        states_d_scores, backprops = _beam_utils.update_beam(self.moves, self.nr_feature, 500,
+                                        states, tokvecs, golds,
+                                        state2vec, vec2scores,
+                                        drop, sgd, losses,
+                                        width=8)
+        backprop_lower = []
+        for i, d_scores in enumerate(states_d_scores):
+            if losses is not None:
+                losses[self.name] += (d_scores**2).sum()
+            ids, bp_vectors, bp_scores = backprops[i]
+            d_vector = bp_scores(d_scores, sgd=sgd)
+            if isinstance(self.model[0].ops, CupyOps) \
+            and not isinstance(ids, state2vec.ops.xp.ndarray):
+                backprop_lower.append((
+                    get_async(cuda_stream, ids),
+                    get_async(cuda_stream, d_vector),
+                    bp_vectors))
+            else:
+                backprop_lower.append((ids, d_vector, bp_vectors))
+        d_tokvecs = self.model[0].ops.allocate(tokvecs.shape)
+        self._make_updates(d_tokvecs, backprop_lower, sgd, cuda_stream)
+        d_tokvecs = self.model[0].ops.unflatten(d_tokvecs, lengths)
+        if USE_FINE_TUNE:
+            bp_my_tokvecs(d_tokvecs, sgd=sgd)
+        return d_tokvecs
 
     def _init_gold_batch(self, whole_docs, whole_golds):
         """Make a square batch, of length equal to the shortest doc. A long
@@ -585,14 +633,10 @@ cdef class Parser:
         xp = get_array_module(d_tokvecs)
         for ids, d_vector, bp_vector in backprops:
             d_state_features = bp_vector(d_vector, sgd=sgd)
-            active_feats = ids * (ids >= 0)
-            active_feats = active_feats.reshape((ids.shape[0], ids.shape[1], 1))
-            if hasattr(xp, 'scatter_add'):
-                xp.scatter_add(d_tokvecs,
-                    ids, d_state_features * active_feats)
-            else:
-                xp.add.at(d_tokvecs,
-                    ids, d_state_features * active_feats)
+            mask = ids >= 0
+            indices = xp.nonzero(mask)
+            self.model[0].ops.scatter_add(d_tokvecs, ids[indices],
+                d_state_features[indices])
 
     @property
     def move_names(self):
@@ -603,7 +647,7 @@ cdef class Parser:
         return names
 
     def get_batch_model(self, batch_size, tokvecs, stream, dropout):
-        lower, upper = self.model
+        _, lower, upper = self.model
         state2vec = precompute_hiddens(batch_size, tokvecs,
                         lower, stream, drop=dropout)
         return state2vec, upper
@@ -693,10 +737,12 @@ cdef class Parser:
 
     def to_disk(self, path, **exclude):
         serializers = {
-            'lower_model': lambda p: p.open('wb').write(
+            'tok2vec_model': lambda p: p.open('wb').write(
                 self.model[0].to_bytes()),
-            'upper_model': lambda p: p.open('wb').write(
+            'lower_model': lambda p: p.open('wb').write(
                 self.model[1].to_bytes()),
+            'upper_model': lambda p: p.open('wb').write(
+                self.model[2].to_bytes()),
             'vocab': lambda p: self.vocab.to_disk(p),
             'moves': lambda p: self.moves.to_disk(p, strings=False),
             'cfg': lambda p: p.open('w').write(json_dumps(self.cfg))
@@ -717,24 +763,29 @@ cdef class Parser:
                 self.model, cfg = self.Model(**self.cfg)
             else:
                 cfg = {}
-            with (path / 'lower_model').open('rb') as file_:
+            with (path / 'tok2vec_model').open('rb') as file_:
                 bytes_data = file_.read()
             self.model[0].from_bytes(bytes_data)
-            with (path / 'upper_model').open('rb') as file_:
+            with (path / 'lower_model').open('rb') as file_:
                 bytes_data = file_.read()
             self.model[1].from_bytes(bytes_data)
+            with (path / 'upper_model').open('rb') as file_:
+                bytes_data = file_.read()
+            self.model[2].from_bytes(bytes_data)
             self.cfg.update(cfg)
         return self
 
     def to_bytes(self, **exclude):
         serializers = OrderedDict((
-            ('lower_model', lambda: self.model[0].to_bytes()),
-            ('upper_model', lambda: self.model[1].to_bytes()),
+            ('tok2vec_model', lambda: self.model[0].to_bytes()),
+            ('lower_model', lambda: self.model[1].to_bytes()),
+            ('upper_model', lambda: self.model[2].to_bytes()),
             ('vocab', lambda: self.vocab.to_bytes()),
             ('moves', lambda: self.moves.to_bytes(strings=False)),
             ('cfg', lambda: ujson.dumps(self.cfg))
         ))
         if 'model' in exclude:
+            exclude['tok2vec_model'] = True
             exclude['lower_model'] = True
             exclude['upper_model'] = True
             exclude.pop('model')
@@ -745,6 +796,7 @@ cdef class Parser:
             ('vocab', lambda b: self.vocab.from_bytes(b)),
             ('moves', lambda b: self.moves.from_bytes(b, strings=False)),
             ('cfg', lambda b: self.cfg.update(ujson.loads(b))),
+            ('tok2vec_model', lambda b: None),
             ('lower_model', lambda b: None),
             ('upper_model', lambda b: None)
         ))
@@ -754,10 +806,12 @@ cdef class Parser:
                 self.model, cfg = self.Model(self.moves.n_moves)
             else:
                 cfg = {}
+            if 'tok2vec_model' in msg:
+                self.model[0].from_bytes(msg['tok2vec_model'])
             if 'lower_model' in msg:
-                self.model[0].from_bytes(msg['lower_model'])
+                self.model[1].from_bytes(msg['lower_model'])
             if 'upper_model' in msg:
-                self.model[1].from_bytes(msg['upper_model'])
+                self.model[2].from_bytes(msg['upper_model'])
             self.cfg.update(cfg)
         return self
 
diff --git a/spacy/syntax/transition_system.pyx b/spacy/syntax/transition_system.pyx
index 27b375bba..d3f64f827 100644
--- a/spacy/syntax/transition_system.pyx
+++ b/spacy/syntax/transition_system.pyx
@@ -107,6 +107,8 @@ cdef class TransitionSystem:
 
     def is_valid(self, StateClass stcls, move_name):
         action = self.lookup_transition(move_name)
+        if action.move == 0:
+            return False
         return action.is_valid(stcls.c, action.label)
 
     cdef int set_valid(self, int* is_valid, const StateC* st) nogil:
diff --git a/spacy/tests/parser/test_neural_parser.py b/spacy/tests/parser/test_neural_parser.py
index 42b55745f..30a6367c8 100644
--- a/spacy/tests/parser/test_neural_parser.py
+++ b/spacy/tests/parser/test_neural_parser.py
@@ -78,3 +78,16 @@ def test_predict_doc_beam(parser, tok2vec, model, doc):
     parser(doc, beam_width=32, beam_density=0.001)
     for word in doc:
         print(word.text, word.head, word.dep_)
+
+
+def test_update_doc_beam(parser, tok2vec, model, doc, gold):
+    parser.model = model
+    tokvecs, bp_tokvecs = tok2vec.begin_update([doc])
+    d_tokvecs = parser.update_beam(([doc], tokvecs), [gold])
+    assert d_tokvecs[0].shape == tokvecs[0].shape
+    def optimize(weights, gradient, key=None):
+        weights -= 0.001 * gradient
+    bp_tokvecs(d_tokvecs, sgd=optimize)
+    assert d_tokvecs[0].sum() == 0.
+
+
diff --git a/spacy/tests/parser/test_nn_beam.py b/spacy/tests/parser/test_nn_beam.py
new file mode 100644
index 000000000..45c85d969
--- /dev/null
+++ b/spacy/tests/parser/test_nn_beam.py
@@ -0,0 +1,87 @@
+from __future__ import unicode_literals
+import pytest
+import numpy
+from thinc.api import layerize
+
+from ...vocab import Vocab
+from ...syntax.arc_eager import ArcEager
+from ...tokens import Doc
+from ...gold import GoldParse
+from ...syntax._beam_utils import ParserBeam, update_beam
+from ...syntax.stateclass import StateClass
+
+
+@pytest.fixture
+def vocab():
+    return Vocab()
+
+@pytest.fixture
+def moves(vocab):
+    aeager = ArcEager(vocab.strings, {})
+    aeager.add_action(2, 'nsubj')
+    aeager.add_action(3, 'dobj')
+    aeager.add_action(2, 'aux')
+    return aeager
+
+
+@pytest.fixture
+def docs(vocab):
+    return [Doc(vocab, words=['Rats', 'bite', 'things'])]
+
+@pytest.fixture
+def states(docs):
+    return [StateClass(doc) for doc in docs]
+
+@pytest.fixture
+def tokvecs(docs, vector_size):
+    output = []
+    for doc in docs:
+        vec = numpy.random.uniform(-0.1, 0.1, (len(doc), vector_size))
+        output.append(numpy.asarray(vec))
+    return output
+
+
+@pytest.fixture
+def golds(docs):
+    return [GoldParse(doc) for doc in docs]
+
+
+@pytest.fixture
+def batch_size(docs):
+    return len(docs)
+
+
+@pytest.fixture
+def beam_width():
+    return 4
+
+
+@pytest.fixture
+def vector_size():
+    return 6
+
+
+@pytest.fixture
+def beam(moves, states, golds, beam_width):
+    return ParserBeam(moves, states, golds, width=beam_width)
+
+@pytest.fixture
+def scores(moves, batch_size, beam_width):
+    return [
+        numpy.asarray(
+            numpy.random.uniform(-0.1, 0.1, (batch_size, moves.n_moves)),
+            dtype='f')
+        for _ in range(batch_size)]
+
+
+def test_create_beam(beam):
+    pass
+
+
+def test_beam_advance(beam, scores):
+    beam.advance(scores)
+
+
+def test_beam_advance_too_few_scores(beam, scores):
+    with pytest.raises(IndexError):
+        beam.advance(scores[:-1])