diff --git a/java/src/com/android/inputmethod/keyboard/LatinKeyboardBaseView.java b/java/src/com/android/inputmethod/keyboard/LatinKeyboardBaseView.java
index b5b6232..b807dd3 100644
--- a/java/src/com/android/inputmethod/keyboard/LatinKeyboardBaseView.java
+++ b/java/src/com/android/inputmethod/keyboard/LatinKeyboardBaseView.java
@@ -37,11 +37,9 @@
 import com.android.inputmethod.keyboard.PointerTracker.DrawingProxy;
 import com.android.inputmethod.keyboard.PointerTracker.TimerProxy;
 import com.android.inputmethod.keyboard.internal.MiniKeyboardBuilder;
-import com.android.inputmethod.keyboard.internal.PointerTrackerQueue;
 import com.android.inputmethod.latin.R;
 import com.android.inputmethod.latin.StaticInnerHandlerWrapper;
 
-import java.util.ArrayList;
 import java.util.WeakHashMap;
 
 /**
@@ -74,11 +72,6 @@
     /** Listener for {@link KeyboardActionListener}. */
     private KeyboardActionListener mKeyboardActionListener;
 
-    private final ArrayList<PointerTracker> mPointerTrackers = new ArrayList<PointerTracker>();
-
-    // TODO: Let the PointerTracker class manage this pointer queue
-    private final PointerTrackerQueue mPointerQueue;
-
     private final boolean mHasDistinctMultitouch;
     private int mOldPointerCount = 1;
     private int mOldKeyIndex;
@@ -251,7 +244,7 @@
                 .hasSystemFeature(PackageManager.FEATURE_TOUCHSCREEN_MULTITOUCH_DISTINCT);
         mKeyRepeatInterval = res.getInteger(R.integer.config_key_repeat_interval);
 
-        mPointerQueue = mHasDistinctMultitouch ? new PointerTrackerQueue() : null;
+        PointerTracker.init(mHasDistinctMultitouch, getContext());
     }
 
     public void startIgnoringDoubleTap() {
@@ -261,9 +254,7 @@
 
     public void setKeyboardActionListener(KeyboardActionListener listener) {
         mKeyboardActionListener = listener;
-        for (PointerTracker tracker : mPointerTrackers) {
-            tracker.setKeyboardActionListener(listener);
-        }
+        PointerTracker.setKeyboardActionListener(listener);
     }
 
     /**
@@ -306,17 +297,15 @@
     @Override
     public void setKeyboard(Keyboard keyboard) {
         if (getKeyboard() != null) {
-            dismissAllKeyPreviews();
+            PointerTracker.dismissAllKeyPreviews();
         }
         // Remove any pending messages, except dismissing preview
         mKeyTimerHandler.cancelKeyTimers();
         super.setKeyboard(keyboard);
         mKeyDetector.setKeyboard(
                 keyboard, -getPaddingLeft(), -getPaddingTop() + mVerticalCorrection);
-        for (PointerTracker tracker : mPointerTrackers) {
-            tracker.setKeyDetector(mKeyDetector);
-        }
         mKeyDetector.setProximityThreshold(keyboard.getMostCommonKeyWidth());
+        PointerTracker.setKeyDetector(mKeyDetector);
         mPopupPanelCache.clear();
     }
 
@@ -345,14 +334,6 @@
         return mKeyDetector.isProximityCorrectionEnabled();
     }
 
-    // TODO: clean up this method.
-    private void dismissAllKeyPreviews() {
-        for (PointerTracker tracker : mPointerTrackers) {
-            tracker.setReleasedKeyGraphics();
-            dismissKeyPreview(tracker);
-        }
-    }
-
     @Override
     public void cancelAllMessages() {
         mKeyTimerHandler.cancelAllMessages();
@@ -451,29 +432,14 @@
     }
 
     private PointerTracker getPointerTracker(final int id) {
-        final ArrayList<PointerTracker> pointers = mPointerTrackers;
-        final KeyboardActionListener listener = mKeyboardActionListener;
-
-        // Create pointer trackers until we can get 'id+1'-th tracker, if needed.
-        for (int i = pointers.size(); i <= id; i++) {
-            final PointerTracker tracker =
-                new PointerTracker(i, getContext(), mKeyTimerHandler, mKeyDetector, this,
-                        mPointerQueue);
-            if (listener != null)
-                tracker.setKeyboardActionListener(listener);
-            pointers.add(tracker);
-        }
-
-        return pointers.get(id);
+        return PointerTracker.getPointerTracker(id, this);
     }
 
     public boolean isInSlidingKeyInput() {
         if (mPopupPanel != null) {
             return true;
-        } else if (mPointerQueue != null) {
-            return mPointerQueue.isInSlidingKeyInput();
         } else {
-            return getPointerTracker(0).isInSlidingKeyInput();
+            return PointerTracker.isAnyInSlidingKeyInput();
         }
     }
 
@@ -499,7 +465,7 @@
         // Gesture detector must be enabled only when mini-keyboard is not on the screen.
         if (mPopupPanel == null && mGestureDetector != null
                 && mGestureDetector.onTouchEvent(me)) {
-            dismissAllKeyPreviews();
+            PointerTracker.dismissAllKeyPreviews();
             mKeyTimerHandler.cancelKeyTimers();
             return true;
         }
diff --git a/java/src/com/android/inputmethod/keyboard/PointerTracker.java b/java/src/com/android/inputmethod/keyboard/PointerTracker.java
index d23fb4a..aa2f3af 100644
--- a/java/src/com/android/inputmethod/keyboard/PointerTracker.java
+++ b/java/src/com/android/inputmethod/keyboard/PointerTracker.java
@@ -25,6 +25,7 @@
 import com.android.inputmethod.latin.R;
 import com.android.inputmethod.latin.SubtypeSwitcher;
 
+import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.List;
 
@@ -77,23 +78,24 @@
         public void cancelKeyTimers();
     }
 
-    public final int mPointerId;
-
+    private static KeyboardSwitcher sKeyboardSwitcher;
+    private static boolean sConfigSlidingKeyInputEnabled;
     // Timing constants
-    private final int mDelayBeforeKeyRepeatStart;
-    private final int mLongPressKeyTimeout;
-    private final int mLongPressShiftKeyTimeout;
+    private static int sDelayBeforeKeyRepeatStart;
+    private static int sLongPressKeyTimeout;
+    private static int sLongPressShiftKeyTimeout;
+    private static int sTouchNoiseThresholdMillis;
+    private static int sTouchNoiseThresholdDistanceSquared;
+
+    private static final List<PointerTracker> sTrackers = new ArrayList<PointerTracker>();
+    private static PointerTrackerQueue sPointerTrackerQueue;
+
+    public final int mPointerId;
 
     private DrawingProxy mDrawingProxy;
     private TimerProxy mTimerProxy;
-    private final PointerTrackerQueue mPointerTrackerQueue;
     private KeyDetector mKeyDetector;
     private KeyboardActionListener mListener = EMPTY_LISTENER;
-    private final KeyboardSwitcher mKeyboardSwitcher;
-    private final boolean mConfigSlidingKeyInputEnabled;
-
-    private final int mTouchNoiseThresholdMillis;
-    private final int mTouchNoiseThresholdDistanceSquared;
 
     private Keyboard mKeyboard;
     private List<Key> mKeys;
@@ -123,7 +125,7 @@
     private boolean mIsRepeatableKey;
 
     // true if this pointer is in sliding key input
-    private boolean mIsInSlidingKeyInput;
+    boolean mIsInSlidingKeyInput;
 
     // true if sliding key is allowed.
     private boolean mIsAllowedSlidingKeyInput;
@@ -135,7 +137,7 @@
     // true if this pointer is in sliding language switch
     private boolean mIsInSlidingLanguageSwitch;
     private int mSpaceKeyIndex;
-    private final SubtypeSwitcher mSubtypeSwitcher;
+    private static SubtypeSwitcher sSubtypeSwitcher;
 
     // Empty {@link KeyboardActionListener}
     private static final KeyboardActionListener EMPTY_LISTENER = new KeyboardActionListener() {
@@ -151,31 +153,72 @@
         public void onCancelInput() {}
     };
 
-    public PointerTracker(int id, Context context, TimerProxy timerProxy, KeyDetector keyDetector,
-            DrawingProxy drawingProxy, PointerTrackerQueue queue) {
-        if (drawingProxy == null || timerProxy == null || keyDetector == null)
-            throw new NullPointerException();
-        mPointerId = id;
-        mDrawingProxy = drawingProxy;
-        mTimerProxy = timerProxy;
-        mPointerTrackerQueue = queue;  // This is null for non-distinct multi-touch device.
-        setKeyDetectorInner(keyDetector);
-        mKeyboardSwitcher = KeyboardSwitcher.getInstance();
+    public static void init(boolean hasDistinctMultitouch, Context context) {
+        if (hasDistinctMultitouch) {
+            sPointerTrackerQueue = new PointerTrackerQueue();
+        } else {
+            sPointerTrackerQueue = null;
+        }
+
         final Resources res = context.getResources();
-        mConfigSlidingKeyInputEnabled = res.getBoolean(R.bool.config_sliding_key_input_enabled);
-        mDelayBeforeKeyRepeatStart = res.getInteger(R.integer.config_delay_before_key_repeat_start);
-        mLongPressKeyTimeout = res.getInteger(R.integer.config_long_press_key_timeout);
-        mLongPressShiftKeyTimeout = res.getInteger(R.integer.config_long_press_shift_key_timeout);
-        mTouchNoiseThresholdMillis = res.getInteger(R.integer.config_touch_noise_threshold_millis);
+        sConfigSlidingKeyInputEnabled = res.getBoolean(R.bool.config_sliding_key_input_enabled);
+        sDelayBeforeKeyRepeatStart = res.getInteger(R.integer.config_delay_before_key_repeat_start);
+        sLongPressKeyTimeout = res.getInteger(R.integer.config_long_press_key_timeout);
+        sLongPressShiftKeyTimeout = res.getInteger(R.integer.config_long_press_shift_key_timeout);
+        sTouchNoiseThresholdMillis = res.getInteger(R.integer.config_touch_noise_threshold_millis);
         final float touchNoiseThresholdDistance = res.getDimension(
                 R.dimen.config_touch_noise_threshold_distance);
-        mTouchNoiseThresholdDistanceSquared = (int)(
+        sTouchNoiseThresholdDistanceSquared = (int)(
                 touchNoiseThresholdDistance * touchNoiseThresholdDistance);
-        mSubtypeSwitcher = SubtypeSwitcher.getInstance();
+        sKeyboardSwitcher = KeyboardSwitcher.getInstance();
+        sSubtypeSwitcher = SubtypeSwitcher.getInstance();
     }
 
-    public void setKeyboardActionListener(KeyboardActionListener listener) {
-        mListener = listener;
+    public static PointerTracker getPointerTracker(final int id, KeyEventHandler handler) {
+        final List<PointerTracker> trackers = sTrackers;
+
+        // Create pointer trackers until we can get 'id+1'-th tracker, if needed.
+        for (int i = trackers.size(); i <= id; i++) {
+            final PointerTracker tracker = new PointerTracker(i, handler);
+            trackers.add(tracker);
+        }
+
+        return trackers.get(id);
+    }
+
+    public static boolean isAnyInSlidingKeyInput() {
+        return sPointerTrackerQueue != null ? sPointerTrackerQueue.isAnyInSlidingKeyInput() : false;
+    }
+
+    public static void setKeyboardActionListener(KeyboardActionListener listener) {
+        for (final PointerTracker tracker : sTrackers) {
+            tracker.mListener = listener;
+        }
+    }
+
+    public static void setKeyDetector(KeyDetector keyDetector) {
+        for (final PointerTracker tracker : sTrackers) {
+            tracker.setKeyDetectorInner(keyDetector);
+            // Mark that keyboard layout has been changed.
+            tracker.mKeyboardLayoutHasBeenChanged = true;
+        }
+    }
+
+    public static void dismissAllKeyPreviews() {
+        for (final PointerTracker tracker : sTrackers) {
+            tracker.setReleasedKeyGraphics();
+            tracker.dismissKeyPreview();
+        }
+    }
+
+    public PointerTracker(int id, KeyEventHandler handler) {
+        if (handler == null)
+            throw new NullPointerException();
+        mPointerId = id;
+        setKeyDetectorInner(handler.getKeyDetector());
+        mListener = handler.getKeyboardActionListener();
+        mDrawingProxy = handler.getDrawingProxy();
+        mTimerProxy = handler.getTimerProxy();
     }
 
     // Returns true if keyboard has been changed by this callback.
@@ -243,14 +286,6 @@
         mKeyQuarterWidthSquared = keyQuarterWidth * keyQuarterWidth;
     }
 
-    public void setKeyDetector(KeyDetector keyDetector) {
-        if (keyDetector == null)
-            throw new NullPointerException();
-        setKeyDetectorInner(keyDetector);
-        // Mark that keyboard layout has been changed.
-        mKeyboardLayoutHasBeenChanged = true;
-    }
-
     public boolean isInSlidingKeyInput() {
         return mIsInSlidingKeyInput;
     }
@@ -365,11 +400,11 @@
         setKeyDetectorInner(handler.getKeyDetector());
         // Naive up-to-down noise filter.
         final long deltaT = eventTime - mUpTime;
-        if (deltaT < mTouchNoiseThresholdMillis) {
+        if (deltaT < sTouchNoiseThresholdMillis) {
             final int dx = x - mLastX;
             final int dy = y - mLastY;
             final int distanceSquared = (dx * dx + dy * dy);
-            if (distanceSquared < mTouchNoiseThresholdDistanceSquared) {
+            if (distanceSquared < sTouchNoiseThresholdDistanceSquared) {
                 if (DEBUG_MODE)
                     Log.w(TAG, "onDownEvent: ignore potential noise: time=" + deltaT
                             + " distance=" + distanceSquared);
@@ -378,7 +413,7 @@
             }
         }
 
-        final PointerTrackerQueue queue = mPointerTrackerQueue;
+        final PointerTrackerQueue queue = sPointerTrackerQueue;
         if (queue != null) {
             if (isOnModifierKey(x, y)) {
                 // Before processing a down event of modifier key, all pointers already being
@@ -394,7 +429,7 @@
         int keyIndex = onDownKey(x, y, eventTime);
         // Sliding key is allowed when 1) enabled by configuration, 2) this pointer starts sliding
         // from modifier key, or 3) this pointer is on mini-keyboard.
-        mIsAllowedSlidingKeyInput = mConfigSlidingKeyInputEnabled || isModifierInternal(keyIndex)
+        mIsAllowedSlidingKeyInput = sConfigSlidingKeyInputEnabled || isModifierInternal(keyIndex)
                 || mKeyDetector instanceof MiniKeyboardKeyDetector;
         mKeyboardLayoutHasBeenChanged = false;
         mKeyAlreadyProcessed = false;
@@ -495,8 +530,8 @@
             else if (isSpaceKey(keyIndex) && !mIsInSlidingLanguageSwitch
                     && mKeyboard instanceof LatinKeyboard) {
                 final LatinKeyboard keyboard = ((LatinKeyboard)mKeyboard);
-                if (mSubtypeSwitcher.useSpacebarLanguageSwitcher()
-                        && mSubtypeSwitcher.getEnabledKeyboardLocaleCount() > 1) {
+                if (sSubtypeSwitcher.useSpacebarLanguageSwitcher()
+                        && sSubtypeSwitcher.getEnabledKeyboardLocaleCount() > 1) {
                     final int diff = x - mKeyX;
                     if (keyboard.shouldTriggerSpacebarSlidingLanguageSwitch(diff)) {
                         // Detect start sliding language switch.
@@ -505,7 +540,7 @@
                         keyboard.updateSpacebarPreviewIcon(diff);
                         // Display spacebar slide language switcher.
                         showKeyPreview(keyIndex);
-                        final PointerTrackerQueue queue = mPointerTrackerQueue;
+                        final PointerTrackerQueue queue = sPointerTrackerQueue;
                         if (queue != null)
                             queue.releaseAllPointersExcept(this, eventTime, true);
                     }
@@ -533,7 +568,7 @@
         if (DEBUG_EVENT)
             printTouchEvent("onUpEvent  :", x, y, eventTime);
 
-        final PointerTrackerQueue queue = mPointerTrackerQueue;
+        final PointerTrackerQueue queue = sPointerTrackerQueue;
         if (queue != null) {
             if (isModifier()) {
                 // Before processing an up event of modifier key, all pointers already being
@@ -600,7 +635,7 @@
         mKeyAlreadyProcessed = true;
         setReleasedKeyGraphics();
         dismissKeyPreview();
-        final PointerTrackerQueue queue = mPointerTrackerQueue;
+        final PointerTrackerQueue queue = sPointerTrackerQueue;
         if (queue != null) {
             queue.remove(this);
         }
@@ -610,7 +645,7 @@
         if (DEBUG_EVENT)
             printTouchEvent("onCancelEvt:", x, y, eventTime);
 
-        final PointerTrackerQueue queue = mPointerTrackerQueue;
+        final PointerTrackerQueue queue = sPointerTrackerQueue;
         if (queue != null) {
             queue.releaseAllPointersExcept(this, eventTime, true);
             queue.remove(this);
@@ -631,7 +666,7 @@
         if (key != null && key.mRepeatable) {
             dismissKeyPreview();
             onRepeatKey(keyIndex);
-            mTimerProxy.startKeyRepeatTimer(mDelayBeforeKeyRepeatStart, keyIndex, this);
+            mTimerProxy.startKeyRepeatTimer(sDelayBeforeKeyRepeatStart, keyIndex, this);
             mIsRepeatableKey = true;
         } else {
             mIsRepeatableKey = false;
@@ -685,16 +720,16 @@
     private void startLongPressTimer(int keyIndex) {
         Key key = getKey(keyIndex);
         if (key.mCode == Keyboard.CODE_SHIFT) {
-            mTimerProxy.startLongPressShiftTimer(mLongPressShiftKeyTimeout, keyIndex, this);
+            mTimerProxy.startLongPressShiftTimer(sLongPressShiftKeyTimeout, keyIndex, this);
         } else if (key.hasUppercaseLetter() && mKeyboard.isManualTemporaryUpperCase()) {
             // We need not start long press timer on the key which has manual temporary upper case
             // code defined and the keyboard is in manual temporary upper case mode.
             return;
-        } else if (mKeyboardSwitcher.isInMomentarySwitchState()) {
+        } else if (sKeyboardSwitcher.isInMomentarySwitchState()) {
             // We use longer timeout for sliding finger input started from the symbols mode key.
-            mTimerProxy.startLongPressTimer(mLongPressKeyTimeout * 3, keyIndex, this);
+            mTimerProxy.startLongPressTimer(sLongPressKeyTimeout * 3, keyIndex, this);
         } else {
-            mTimerProxy.startLongPressTimer(mLongPressKeyTimeout, keyIndex, this);
+            mTimerProxy.startLongPressTimer(sLongPressKeyTimeout, keyIndex, this);
         }
     }
 
diff --git a/java/src/com/android/inputmethod/keyboard/internal/PointerTrackerQueue.java b/java/src/com/android/inputmethod/keyboard/internal/PointerTrackerQueue.java
index f87cd86..545b27f 100644
--- a/java/src/com/android/inputmethod/keyboard/internal/PointerTrackerQueue.java
+++ b/java/src/com/android/inputmethod/keyboard/internal/PointerTrackerQueue.java
@@ -63,7 +63,7 @@
         mQueue.remove(tracker);
     }
 
-    public boolean isInSlidingKeyInput() {
+    public boolean isAnyInSlidingKeyInput() {
         for (final PointerTracker tracker : mQueue) {
             if (tracker.isInSlidingKeyInput())
                 return true;
diff --git a/native/Android.mk b/native/Android.mk
index d9f4f1d..c221913 100644
--- a/native/Android.mk
+++ b/native/Android.mk
@@ -19,6 +19,7 @@
     src/unigram_dictionary.cpp
 
 #FLAG_DBG := true
+#FLAG_DO_PROFILE := true
 
 TARGETING_UNBUNDLED_FROYO := true
 
@@ -39,10 +40,15 @@
 
 LOCAL_MODULE_TAGS := user
 
+ifeq ($(FLAG_DO_PROFILE), true)
+    $(warning Making profiling version of native library)
+    LOCAL_CFLAGS += -DFLAG_DO_PROFILE
+else # FLAG_DO_PROFILE
 ifeq ($(FLAG_DBG), true)
     $(warning Making debug version of native library)
     LOCAL_CFLAGS += -DFLAG_DBG
     LOCAL_SHARED_LIBRARIES := libcutils libutils
-endif
+endif # FLAG_DBG
+endif # FLAG_DO_PROFILE
 
 include $(BUILD_SHARED_LIBRARY)
diff --git a/native/src/binary_format.h b/native/src/binary_format.h
new file mode 100644
index 0000000..e9f108e
--- /dev/null
+++ b/native/src/binary_format.h
@@ -0,0 +1,209 @@
+/*
+ * Copyright (C) 2011 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef LATINIME_BINARY_FORMAT_H
+#define LATINIME_BINARY_FORMAT_H
+
+namespace latinime {
+
+class BinaryFormat {
+private:
+    const static int32_t MINIMAL_ONE_BYTE_CHARACTER_VALUE = 0x20;
+    const static int32_t CHARACTER_ARRAY_TERMINATOR = 0x1F;
+    const static int MULTIPLE_BYTE_CHARACTER_ADDITIONAL_SIZE = 2;
+
+public:
+    static int getGroupCountAndForwardPointer(const uint8_t* const dict, int* pos);
+    static uint8_t getFlagsAndForwardPointer(const uint8_t* const dict, int* pos);
+    static int32_t getCharCodeAndForwardPointer(const uint8_t* const dict, int* pos);
+    static int readFrequencyWithoutMovingPointer(const uint8_t* const dict, const int pos);
+    static int skipOtherCharacters(const uint8_t* const dict, const int pos);
+    static int skipAttributes(const uint8_t* const dict, const int pos);
+    static int skipChildrenPosition(const uint8_t flags, const int pos);
+    static int skipFrequency(const uint8_t flags, const int pos);
+    static int skipAllAttributes(const uint8_t* const dict, const uint8_t flags, const int pos);
+    static int skipChildrenPosAndAttributes(const uint8_t* const dict, const uint8_t flags,
+            const int pos);
+    static int readChildrenPosition(const uint8_t* const dict, const uint8_t flags, const int pos);
+    static bool hasChildrenInFlags(const uint8_t flags);
+    static int getAttributeAddressAndForwardPointer(const uint8_t* const dict, const uint8_t flags,
+            int *pos);
+};
+
+inline int BinaryFormat::getGroupCountAndForwardPointer(const uint8_t* const dict, int* pos) {
+    return dict[(*pos)++];
+}
+
+inline uint8_t BinaryFormat::getFlagsAndForwardPointer(const uint8_t* const dict, int* pos) {
+    return dict[(*pos)++];
+}
+
+inline int32_t BinaryFormat::getCharCodeAndForwardPointer(const uint8_t* const dict, int* pos) {
+    const int origin = *pos;
+    const int32_t character = dict[origin];
+    if (character < MINIMAL_ONE_BYTE_CHARACTER_VALUE) {
+        if (character == CHARACTER_ARRAY_TERMINATOR) {
+            *pos = origin + 1;
+            return NOT_A_CHARACTER;
+        } else {
+            *pos = origin + 3;
+            const int32_t char_1 = character << 16;
+            const int32_t char_2 = char_1 + (dict[origin + 1] << 8);
+            return char_2 + dict[origin + 2];
+        }
+    } else {
+        *pos = origin + 1;
+        return character;
+    }
+}
+
+inline int BinaryFormat::readFrequencyWithoutMovingPointer(const uint8_t* const dict,
+        const int pos) {
+    return dict[pos];
+}
+
+inline int BinaryFormat::skipOtherCharacters(const uint8_t* const dict, const int pos) {
+    int currentPos = pos;
+    int32_t character = dict[currentPos++];
+    while (CHARACTER_ARRAY_TERMINATOR != character) {
+        if (character < MINIMAL_ONE_BYTE_CHARACTER_VALUE) {
+            currentPos += MULTIPLE_BYTE_CHARACTER_ADDITIONAL_SIZE;
+        }
+        character = dict[currentPos++];
+    }
+    return currentPos;
+}
+
+static inline int attributeAddressSize(const uint8_t flags) {
+    static const int ATTRIBUTE_ADDRESS_SHIFT = 4;
+    return (flags & UnigramDictionary::MASK_ATTRIBUTE_ADDRESS_TYPE) >> ATTRIBUTE_ADDRESS_SHIFT;
+    /* Note: this is a value-dependant optimization of what may probably be
+       more readably written this way:
+       switch (flags * UnigramDictionary::MASK_ATTRIBUTE_ADDRESS_TYPE) {
+       case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_ONEBYTE: return 1;
+       case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_TWOBYTES: return 2;
+       case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_THREEBYTE: return 3;
+       default: return 0;
+       }
+    */
+}
+
+inline int BinaryFormat::skipAttributes(const uint8_t* const dict, const int pos) {
+    int currentPos = pos;
+    uint8_t flags = getFlagsAndForwardPointer(dict, &currentPos);
+    while (flags & UnigramDictionary::FLAG_ATTRIBUTE_HAS_NEXT) {
+        currentPos += attributeAddressSize(flags);
+        flags = getFlagsAndForwardPointer(dict, &currentPos);
+    }
+    currentPos += attributeAddressSize(flags);
+    return currentPos;
+}
+
+static inline int childrenAddressSize(const uint8_t flags) {
+    static const int CHILDREN_ADDRESS_SHIFT = 6;
+    return (UnigramDictionary::MASK_GROUP_ADDRESS_TYPE & flags) >> CHILDREN_ADDRESS_SHIFT;
+    /* See the note in attributeAddressSize. The same applies here */
+}
+
+inline int BinaryFormat::skipChildrenPosition(const uint8_t flags, const int pos) {
+    return pos + childrenAddressSize(flags);
+}
+
+inline int BinaryFormat::skipFrequency(const uint8_t flags, const int pos) {
+    return UnigramDictionary::FLAG_IS_TERMINAL & flags ? pos + 1 : pos;
+}
+
+inline int BinaryFormat::skipAllAttributes(const uint8_t* const dict, const uint8_t flags,
+        const int pos) {
+    // This function skips all attributes. The format makes provision for future extension
+    // with other attributes (notably shortcuts) but for the time being, bigrams are the
+    // only attributes that may be found in a character group, so we only look at bigrams
+    // in this version.
+    if (UnigramDictionary::FLAG_HAS_BIGRAMS & flags) {
+        return skipAttributes(dict, pos);
+    } else {
+        return pos;
+    }
+}
+
+inline int BinaryFormat::skipChildrenPosAndAttributes(const uint8_t* const dict,
+        const uint8_t flags, const int pos) {
+    int currentPos = pos;
+    currentPos = skipChildrenPosition(flags, currentPos);
+    currentPos = skipAllAttributes(dict, flags, currentPos);
+    return currentPos;
+}
+
+inline int BinaryFormat::readChildrenPosition(const uint8_t* const dict, const uint8_t flags,
+        const int pos) {
+    int offset = 0;
+    switch (UnigramDictionary::MASK_GROUP_ADDRESS_TYPE & flags) {
+        case UnigramDictionary::FLAG_GROUP_ADDRESS_TYPE_ONEBYTE:
+            offset = dict[pos];
+            break;
+        case UnigramDictionary::FLAG_GROUP_ADDRESS_TYPE_TWOBYTES:
+            offset = dict[pos] << 8;
+            offset += dict[pos + 1];
+            break;
+        case UnigramDictionary::FLAG_GROUP_ADDRESS_TYPE_THREEBYTES:
+            offset = dict[pos] << 16;
+            offset += dict[pos + 1] << 8;
+            offset += dict[pos + 2];
+            break;
+        default:
+            // If we come here, it means we asked for the children of a word with
+            // no children.
+            return -1;
+    }
+    return pos + offset;
+}
+
+inline bool BinaryFormat::hasChildrenInFlags(const uint8_t flags) {
+    return (UnigramDictionary::FLAG_GROUP_ADDRESS_TYPE_NOADDRESS
+            != (UnigramDictionary::MASK_GROUP_ADDRESS_TYPE & flags));
+}
+
+inline int BinaryFormat::getAttributeAddressAndForwardPointer(const uint8_t* const dict,
+        const uint8_t flags, int *pos) {
+    int offset = 0;
+    const int origin = *pos;
+    switch (UnigramDictionary::MASK_ATTRIBUTE_ADDRESS_TYPE & flags) {
+        case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_ONEBYTE:
+            offset = dict[origin];
+            *pos = origin + 1;
+            break;
+        case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_TWOBYTES:
+            offset = dict[origin] << 8;
+            offset += dict[origin + 1];
+            *pos = origin + 2;
+            break;
+        case UnigramDictionary::FLAG_ATTRIBUTE_ADDRESS_TYPE_THREEBYTES:
+            offset = dict[origin] << 16;
+            offset += dict[origin + 1] << 8;
+            offset += dict[origin + 2];
+            *pos = origin + 3;
+            break;
+    }
+    if (UnigramDictionary::FLAG_ATTRIBUTE_OFFSET_NEGATIVE & flags) {
+        return origin - offset;
+    } else {
+        return origin + offset;
+    }
+}
+
+} // namespace latinime
+
+#endif // LATINIME_BINARY_FORMAT_H
diff --git a/native/src/defines.h b/native/src/defines.h
index 0a32405..a516190 100644
--- a/native/src/defines.h
+++ b/native/src/defines.h
@@ -18,18 +18,7 @@
 #ifndef LATINIME_DEFINES_H
 #define LATINIME_DEFINES_H
 
-#ifdef FLAG_DBG
-#include <cutils/log.h>
-#ifndef LOG_TAG
-#define LOG_TAG "LatinIME: "
-#endif
-#define DEBUG_DICT true
-#define DEBUG_DICT_FULL false
-#define DEBUG_SHOW_FOUND_WORD DEBUG_DICT_FULL
-#define DEBUG_NODE DEBUG_DICT_FULL
-#define DEBUG_TRACE DEBUG_DICT_FULL
-#define DEBUG_PROXIMITY_INFO true
-
+#ifdef FLAG_DO_PROFILE
 // Profiler
 #include <time.h>
 #define PROF_BUF_SIZE 100
@@ -76,16 +65,7 @@
     }
 }
 
-#else // FLAG_DBG
-#define LOGE(fmt, ...)
-#define LOGI(fmt, ...)
-#define DEBUG_DICT false
-#define DEBUG_DICT_FULL false
-#define DEBUG_SHOW_FOUND_WORD false
-#define DEBUG_NODE false
-#define DEBUG_TRACE false
-#define DEBUG_PROXIMITY_INFO false
-
+#else // FLAG_DO_PROFILE
 #define PROF_BUF_SIZE 0
 #define PROF_RESET
 #define PROF_COUNT(prof_buf_id)
@@ -97,6 +77,30 @@
 #define PROF_CLOCKOUT(prof_buf_id)
 #define PROF_OUTALL
 
+#endif // FLAG_DO_PROFILE
+
+#ifdef FLAG_DBG
+#include <cutils/log.h>
+#ifndef LOG_TAG
+#define LOG_TAG "LatinIME: "
+#endif
+#define DEBUG_DICT true
+#define DEBUG_DICT_FULL false
+#define DEBUG_SHOW_FOUND_WORD DEBUG_DICT_FULL
+#define DEBUG_NODE DEBUG_DICT_FULL
+#define DEBUG_TRACE DEBUG_DICT_FULL
+#define DEBUG_PROXIMITY_INFO true
+
+#else // FLAG_DBG
+#define LOGE(fmt, ...)
+#define LOGI(fmt, ...)
+#define DEBUG_DICT false
+#define DEBUG_DICT_FULL false
+#define DEBUG_SHOW_FOUND_WORD false
+#define DEBUG_NODE false
+#define DEBUG_TRACE false
+#define DEBUG_PROXIMITY_INFO false
+
 #endif // FLAG_DBG
 
 #ifndef U_SHORT_MAX
@@ -126,8 +130,11 @@
 #define FLAG_BIGRAM_FREQ 0x7F
 
 #define DICTIONARY_VERSION_MIN 200
+// TODO: remove this constant when the switch to the new dict format is over
 #define DICTIONARY_HEADER_SIZE 2
+#define NEW_DICTIONARY_HEADER_SIZE 5
 #define NOT_VALID_WORD -99
+#define NOT_A_CHARACTER -1
 
 #define KEYCODE_SPACE ' '
 
diff --git a/native/src/unigram_dictionary.cpp b/native/src/unigram_dictionary.cpp
index 42951bf..698584e 100644
--- a/native/src/unigram_dictionary.cpp
+++ b/native/src/unigram_dictionary.cpp
@@ -25,6 +25,10 @@
 #include "dictionary.h"
 #include "unigram_dictionary.h"
 
+#ifdef NEW_DICTIONARY_FORMAT
+#include "binary_format.h"
+#endif // NEW_DICTIONARY_FORMAT
+
 namespace latinime {
 
 const UnigramDictionary::digraph_t UnigramDictionary::GERMAN_UMLAUT_DIGRAPHS[] =
@@ -36,11 +40,20 @@
 UnigramDictionary::UnigramDictionary(const uint8_t* const streamStart, int typedLetterMultiplier,
         int fullWordMultiplier, int maxWordLength, int maxWords, int maxProximityChars,
         const bool isLatestDictVersion)
+#ifndef NEW_DICTIONARY_FORMAT
     : DICT_ROOT(streamStart),
+#else // NEW_DICTIONARY_FORMAT
+    : DICT_ROOT(streamStart + NEW_DICTIONARY_HEADER_SIZE),
+#endif // NEW_DICTIONARY_FORMAT
     MAX_WORD_LENGTH(maxWordLength), MAX_WORDS(maxWords),
     MAX_PROXIMITY_CHARS(maxProximityChars), IS_LATEST_DICT_VERSION(isLatestDictVersion),
     TYPED_LETTER_MULTIPLIER(typedLetterMultiplier), FULL_WORD_MULTIPLIER(fullWordMultiplier),
+#ifndef NEW_DICTIONARY_FORMAT
     ROOT_POS(isLatestDictVersion ? DICTIONARY_HEADER_SIZE : 0),
+#else // NEW_DICTIONARY_FORMAT
+      // TODO : remove this variable.
+    ROOT_POS(0),
+#endif // NEW_DICTIONARY_FORMAT
     BYTES_IN_ONE_CHAR(MAX_PROXIMITY_CHARS * sizeof(*mInputCodes)),
     MAX_UMLAUT_SEARCH_DEPTH(DEFAULT_MAX_UMLAUT_SEARCH_DEPTH) {
     if (DEBUG_DICT) {
@@ -722,8 +735,6 @@
 }
 
 #ifndef NEW_DICTIONARY_FORMAT
-// TODO: Don't forget to bring inline functions back to over where they are used.
-
 // The following functions will be entirely replaced with new implementations.
 void UnigramDictionary::getWordsOld(const int initialPos, const int inputLength, const int skipPos,
         const int excessivePos, const int transposedPos,int *nextLetters,
@@ -999,10 +1010,241 @@
 
 #else // NEW_DICTIONARY_FORMAT
 
+// Wrapper for getMostFrequentWordLikeInner, which matches it to the previous
+// interface.
+inline int UnigramDictionary::getMostFrequentWordLike(const int startInputIndex,
+        const int inputLength, unsigned short *word) {
+    uint16_t inWord[inputLength];
+
+    for (int i = 0; i < inputLength; ++i) {
+        inWord[i] = *getInputCharsAt(startInputIndex + i);
+    }
+    return getMostFrequentWordLikeInner(inWord, inputLength, word);
+}
+
+// This function will take the position of a character array within a CharGroup,
+// and check it actually like-matches the word in inWord starting at startInputIndex,
+// that is, it matches it with case and accents squashed.
+// The function returns true if there was a full match, false otherwise.
+// The function will copy on-the-fly the characters in the CharGroup to outNewWord.
+// It will also place the end position of the array in outPos; in outInputIndex,
+// it will place the index of the first char AFTER the match if there was a match,
+// and the initial position if there was not. It makes sense because if there was
+// a match we want to continue searching, but if there was not, we want to go to
+// the next CharGroup.
+// In and out parameters may point to the same location. This function takes care
+// not to use any input parameters after it wrote into its outputs.
+static inline bool testCharGroupForContinuedLikeness(const uint8_t flags,
+        const uint8_t* const root, const int startPos,
+        const uint16_t* const inWord, const int startInputIndex,
+        int32_t* outNewWord, int* outInputIndex, int* outPos) {
+    const bool hasMultipleChars = (0 != (UnigramDictionary::FLAG_HAS_MULTIPLE_CHARS & flags));
+    int pos = startPos;
+    int32_t character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+    int32_t baseChar = toBaseLowerCase(character);
+    const uint16_t wChar = toBaseLowerCase(inWord[startInputIndex]);
+
+    if (baseChar != wChar) {
+        *outPos = hasMultipleChars ? BinaryFormat::skipOtherCharacters(root, pos) : pos;
+        *outInputIndex = startInputIndex;
+        return false;
+    }
+    int inputIndex = startInputIndex;
+    outNewWord[inputIndex] = character;
+    if (hasMultipleChars) {
+        character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+        while (NOT_A_CHARACTER != character) {
+            baseChar = toBaseLowerCase(character);
+            if (toBaseLowerCase(inWord[++inputIndex]) != baseChar) {
+                *outPos = BinaryFormat::skipOtherCharacters(root, pos);
+                *outInputIndex = startInputIndex;
+                return false;
+            }
+            outNewWord[inputIndex] = character;
+            character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+        }
+    }
+    *outInputIndex = inputIndex + 1;
+    *outPos = pos;
+    return true;
+}
+
+// This function is invoked when a word like the word searched for is found.
+// It will compare the frequency to the max frequency, and if greater, will
+// copy the word into the output buffer. In output value maxFreq, it will
+// write the new maximum frequency if it changed.
+static inline void onTerminalWordLike(const int freq, int32_t* newWord, const int length,
+        short unsigned int* outWord, int* maxFreq) {
+    if (freq > *maxFreq) {
+        for (int q = 0; q < length; ++q)
+            outWord[q] = newWord[q];
+        outWord[length] = 0;
+        *maxFreq = freq;
+    }
+}
+
+// Will find the highest frequency of the words like the one passed as an argument,
+// that is, everything that only differs by case/accents.
+int UnigramDictionary::getMostFrequentWordLikeInner(const uint16_t * const inWord,
+        const int length, short unsigned int* outWord) {
+    int32_t newWord[MAX_WORD_LENGTH_INTERNAL];
+    int depth = 0;
+    int maxFreq = -1;
+    const uint8_t* const root = DICT_ROOT;
+
+    mStackChildCount[0] = root[0];
+    mStackInputIndex[0] = 0;
+    mStackSiblingPos[0] = 1;
+    while (depth >= 0) {
+        const int charGroupCount = mStackChildCount[depth];
+        int pos = mStackSiblingPos[depth];
+        for (int charGroupIndex = charGroupCount - 1; charGroupIndex >= 0; --charGroupIndex) {
+            int inputIndex = mStackInputIndex[depth];
+            const uint8_t flags = BinaryFormat::getFlagsAndForwardPointer(root, &pos);
+            // Test whether all chars in this group match with the word we are searching for. If so,
+            // we want to traverse its children (or if the length match, evaluate its frequency).
+            // Note that this function will output the position regardless, but will only write
+            // into inputIndex if there is a match.
+            const bool isAlike = testCharGroupForContinuedLikeness(flags, root, pos, inWord,
+                    inputIndex, newWord, &inputIndex, &pos);
+            if (isAlike && (FLAG_IS_TERMINAL & flags) && (inputIndex == length)) {
+                const int frequency = BinaryFormat::readFrequencyWithoutMovingPointer(root, pos);
+                onTerminalWordLike(frequency, newWord, inputIndex, outWord, &maxFreq);
+            }
+            pos = BinaryFormat::skipFrequency(flags, pos);
+            const int siblingPos = BinaryFormat::skipChildrenPosAndAttributes(root, flags, pos);
+            const int childrenNodePos = BinaryFormat::readChildrenPosition(root, flags, pos);
+            // If we had a match and the word has children, we want to traverse them. We don't have
+            // to traverse words longer than the one we are searching for, since they will not match
+            // anyway, so don't traverse unless inputIndex < length.
+            if (isAlike && (-1 != childrenNodePos) && (inputIndex < length)) {
+                // Save position for this depth, to get back to this once children are done
+                mStackChildCount[depth] = charGroupIndex;
+                mStackSiblingPos[depth] = siblingPos;
+                // Prepare stack values for next depth
+                ++depth;
+                int childrenPos = childrenNodePos;
+                mStackChildCount[depth] =
+                        BinaryFormat::getGroupCountAndForwardPointer(root, &childrenPos);
+                mStackSiblingPos[depth] = childrenPos;
+                mStackInputIndex[depth] = inputIndex;
+                pos = childrenPos;
+                // Go to the next depth level.
+                ++depth;
+                break;
+            } else {
+                // No match, or no children, or word too long to ever match: go the next sibling.
+                pos = siblingPos;
+            }
+        }
+        --depth;
+    }
+    return maxFreq;
+}
+
+// This function gets the frequency of the exact matching word in the dictionary.
+// If no match is found, it returns -1.
+int UnigramDictionary::getFrequency(const uint16_t* const inWord, const int length) const {
+    int pos = 0;
+    int wordPos = 0;
+    const uint8_t* const root = DICT_ROOT;
+
+    while (true) {
+        // If we already traversed the tree further than the word is long, there means
+        // there was no match (or we would have found it).
+        if (wordPos > length) return -1;
+        int charGroupCount = BinaryFormat::getGroupCountAndForwardPointer(root, &pos);
+        const uint16_t wChar = inWord[wordPos];
+        while (true) {
+            // If there are no more character groups in this node, it means we could not
+            // find a matching character for this depth, therefore there is no match.
+            if (0 >= charGroupCount) return -1;
+            const uint8_t flags = BinaryFormat::getFlagsAndForwardPointer(root, &pos);
+            int32_t character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+            if (character == wChar) {
+                // This is the correct node. Only one character group may start with the same
+                // char within a node, so either we found our match in this node, or there is
+                // no match and we can return -1. So we will check all the characters in this
+                // character group indeed does match.
+                if (FLAG_HAS_MULTIPLE_CHARS & flags) {
+                    character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+                    while (NOT_A_CHARACTER != character) {
+                        ++wordPos;
+                        // If we shoot the length of the word we search for, or if we find a single
+                        // character that does not match, as explained above, it means the word is
+                        // not in the dictionary (by virtue of this chargroup being the only one to
+                        // match the word on the first character, but not matching the whole word).
+                        if (wordPos > length) return -1;
+                        if (inWord[wordPos] != character) return -1;
+                        character = BinaryFormat::getCharCodeAndForwardPointer(root, &pos);
+                    }
+                }
+                // If we come here we know that so far, we do match. Either we are on a terminal
+                // and we match the length, in which case we found it, or we traverse children.
+                // If we don't match the length AND don't have children, then a word in the
+                // dictionary fully matches a prefix of the searched word but not the full word.
+                ++wordPos;
+                if (FLAG_IS_TERMINAL & flags) {
+                    if (wordPos == length) {
+                        return BinaryFormat::readFrequencyWithoutMovingPointer(root, pos);
+                    }
+                    pos = BinaryFormat::skipFrequency(FLAG_IS_TERMINAL, pos);
+                }
+                if (FLAG_GROUP_ADDRESS_TYPE_NOADDRESS == (MASK_GROUP_ADDRESS_TYPE & flags))
+                    return -1;
+                // We have children and we are still shorter than the word we are searching for, so
+                // we need to traverse children. Put the pointer on the children position, and
+                // break
+                pos = BinaryFormat::readChildrenPosition(root, flags, pos);
+                break;
+            } else {
+                // This chargroup does not match, so skip the remaining part and go to the next.
+                if (FLAG_HAS_MULTIPLE_CHARS & flags) {
+                    pos = BinaryFormat::skipOtherCharacters(root, pos);
+                }
+                pos = BinaryFormat::skipFrequency(flags, pos);
+                pos = BinaryFormat::skipChildrenPosAndAttributes(root, flags, pos);
+            }
+            --charGroupCount;
+        }
+    }
+}
+
+bool UnigramDictionary::isValidWord(const uint16_t* const inWord, const int length) const {
+    return -1 != getFrequency(inWord, length);
+}
+
+int UnigramDictionary::getBigrams(unsigned short *word, int length, int *codes, int codesSize,
+        unsigned short *outWords, int *frequencies, int maxWordLength, int maxBigrams,
+        int maxAlternatives) {
+    // TODO: add implementation.
+    return 0;
+}
+
+// TODO: remove this function.
+int UnigramDictionary::getBigramPosition(int pos, unsigned short *word, int offset,
+        int length) const {
+    return -1;
+}
+
+// ProcessCurrentNode returns a boolean telling whether to traverse children nodes or not.
+// If the return value is false, then the caller should read in the output "nextSiblingPosition"
+// to find out the address of the next sibling node and pass it to a new call of processCurrentNode.
+// It is worthy to note that when false is returned, the output values other than
+// nextSiblingPosition are undefined.
+// If the return value is true, then the caller must proceed to traverse the children of this
+// node. processCurrentNode will output the information about the children: their count in
+// newCount, their position in newChildrenPosition, the traverseAllNodes flag in
+// newTraverseAllNodes, the match weight into newMatchRate, the input index into newInputIndex, the
+// diffs into newDiffs, the sibling position in nextSiblingPosition, and the output index into
+// newOutputIndex. Please also note the following caveat: processCurrentNode does not know when
+// there aren't any more nodes at this level, it merely returns the address of the first byte after
+// the current node in nextSiblingPosition. Thus, the caller must keep count of the nodes at any
+// given level, as output into newCount when traversing this level's parent.
 inline bool UnigramDictionary::processCurrentNode(const int initialPos, const int initialDepth,
         const int maxDepth, const bool initialTraverseAllNodes, int matchWeight, int inputIndex,
         const int initialDiffs, const int skipPos, const int excessivePos, const int transposedPos,
-        int *nextLetters, const int nextLettersSize, int *newCount, int *newChildPosition,
+        int *nextLetters, const int nextLettersSize, int *newCount, int *newChildrenPosition,
         bool *newTraverseAllNodes, int *newMatchRate, int *newInputIndex, int *newDiffs,
         int *nextSiblingPosition, int *newOutputIndex) {
     if (DEBUG_DICT) {
@@ -1012,84 +1254,187 @@
         if (transposedPos >= 0) ++inputCount;
         assert(inputCount <= 1);
     }
-    unsigned short c;
-    int childPosition;
-    bool terminal;
-    int freq;
-    bool isSameAsUserTypedLength = false;
-
     int pos = initialPos;
     int depth = initialDepth;
     int traverseAllNodes = initialTraverseAllNodes;
     int diffs = initialDiffs;
 
-    const uint8_t flags = 0; // No flags for now
+    // Flags contain the following information:
+    // - Address type (MASK_GROUP_ADDRESS_TYPE) on two bits:
+    //   - FLAG_GROUP_ADDRESS_TYPE_{ONE,TWO,THREE}_BYTES means there are children and their address
+    //     is on the specified number of bytes.
+    //   - FLAG_GROUP_ADDRESS_TYPE_NOADDRESS means there are no children, and therefore no address.
+    // - FLAG_HAS_MULTIPLE_CHARS: whether this node has multiple char or not.
+    // - FLAG_IS_TERMINAL: whether this node is a terminal or not (it may still have children)
+    // - FLAG_HAS_BIGRAMS: whether this node has bigrams or not
+    const uint8_t flags = BinaryFormat::getFlagsAndForwardPointer(DICT_ROOT, &pos);
+    const bool hasMultipleChars = (0 != (FLAG_HAS_MULTIPLE_CHARS & flags));
 
-    if (excessivePos == depth && inputIndex < mInputLength - 1) ++inputIndex;
+    // This gets only ONE character from the stream. Next there will be:
+    // if FLAG_HAS_MULTIPLE CHARS: the other characters of the same node
+    // else if FLAG_IS_TERMINAL: the frequency
+    // else if MASK_GROUP_ADDRESS_TYPE is not NONE: the children address
+    // Note that you can't have a node that both is not a terminal and has no children.
+    int32_t c = BinaryFormat::getCharCodeAndForwardPointer(DICT_ROOT, &pos);
+    assert(NOT_A_CHARACTER != c);
 
-    *nextSiblingPosition = Dictionary::setDictionaryValues(DICT_ROOT, IS_LATEST_DICT_VERSION, pos,
-            &c, &childPosition, &terminal, &freq);
-    *newOutputIndex = depth + 1;
+    // We are going to loop through each character and make it look like it's a different
+    // node each time. To do that, we will process characters in this node in order until
+    // we find the character terminator. This is signalled by getCharCode* returning
+    // NOT_A_CHARACTER.
+    // As a special case, if there is only one character in this node, we must not read the
+    // next bytes so we will simulate the NOT_A_CHARACTER return by testing the flags.
+    // This way, each loop run will look like a "virtual node".
+    do {
+        // We prefetch the next char. If 'c' is the last char of this node, we will have
+        // NOT_A_CHARACTER in the next char. From this we can decide whether this virtual node
+        // should behave as a terminal or not and whether we have children.
+        const int32_t nextc = hasMultipleChars
+                ? BinaryFormat::getCharCodeAndForwardPointer(DICT_ROOT, &pos) : NOT_A_CHARACTER;
+        const bool isLastChar = (NOT_A_CHARACTER == nextc);
+        // If there are more chars in this nodes, then this virtual node is not a terminal.
+        // If we are on the last char, this virtual node is a terminal if this node is.
+        const bool isTerminal = isLastChar && (0 != (FLAG_IS_TERMINAL & flags));
+        // If there are more chars in this node, then this virtual node has children.
+        // If we are on the last char, this virtual node has children if this node has.
+        const bool hasChildren = (!isLastChar) || BinaryFormat::hasChildrenInFlags(flags);
 
-    const bool needsToTraverseChildrenNodes = childPosition != 0;
+        // This has to be done for each virtual char (this forwards the "inputIndex" which
+        // is the index in the user-inputted chars, as read by getInputCharsAt.
+        if (excessivePos == depth && inputIndex < mInputLength - 1) ++inputIndex;
+        if (traverseAllNodes || needsToSkipCurrentNode(c, inputIndex, skipPos, depth)) {
+            mWord[depth] = c;
+            if (traverseAllNodes && isTerminal) {
+                // The frequency should be here, because we come here only if this is actually
+                // a terminal node, and we are on its last char.
+                const int freq = BinaryFormat::readFrequencyWithoutMovingPointer(DICT_ROOT, pos);
+                onTerminal(mWord, depth, DICT_ROOT, flags, pos, inputIndex, matchWeight, skipPos,
+                           excessivePos, transposedPos, freq, false, nextLetters, nextLettersSize);
+            }
+            if (!hasChildren) {
+                // If we don't have children here, that means we finished processing all
+                // characters of this node (we are on the last virtual node), AND we are in
+                // traverseAllNodes mode, which means we are searching for *completions*. We
+                // should skip the frequency if we have a terminal, and report the position
+                // of the next sibling. We don't have to return other values because we are
+                // returning false, as in "don't traverse children".
+                if (isTerminal) pos = BinaryFormat::skipFrequency(flags, pos);
+                *nextSiblingPosition =
+                        BinaryFormat::skipChildrenPosAndAttributes(DICT_ROOT, flags, pos);
+                return false;
+            }
+        } else {
+            const int *currentChars = getInputCharsAt(inputIndex);
 
-    // If we are only doing traverseAllNodes, no need to look at the typed characters.
-    if (traverseAllNodes || needsToSkipCurrentNode(c, inputIndex, skipPos, depth)) {
-        mWord[depth] = c;
-        if (traverseAllNodes && terminal) {
-            onTerminal(mWord, depth, DICT_ROOT, flags, pos, inputIndex, matchWeight, skipPos,
-                       excessivePos, transposedPos, freq, false, nextLetters, nextLettersSize);
+            if (transposedPos >= 0) {
+                if (inputIndex == transposedPos) currentChars += MAX_PROXIMITY_CHARS;
+                if (inputIndex == (transposedPos + 1)) currentChars -= MAX_PROXIMITY_CHARS;
+            }
+
+            const int matchedProximityCharId = getMatchedProximityId(currentChars, c, skipPos,
+                    excessivePos, transposedPos);
+            if (UNRELATED_CHAR == matchedProximityCharId) {
+                // We found that this is an unrelated character, so we should give up traversing
+                // this node and its children entirely.
+                // However we may not be on the last virtual node yet so we skip the remaining
+                // characters in this node, the frequency if it's there, read the next sibling
+                // position to output it, then return false.
+                // We don't have to output other values because we return false, as in
+                // "don't traverse children".
+                if (!isLastChar) {
+                    pos = BinaryFormat::skipOtherCharacters(DICT_ROOT, pos);
+                }
+                pos = BinaryFormat::skipFrequency(flags, pos);
+                *nextSiblingPosition =
+                        BinaryFormat::skipChildrenPosAndAttributes(DICT_ROOT, flags, pos);
+                return false;
+            }
+            mWord[depth] = c;
+            // If inputIndex is greater than mInputLength, that means there is no
+            // proximity chars. So, we don't need to check proximity.
+            if (SAME_OR_ACCENTED_OR_CAPITALIZED_CHAR == matchedProximityCharId) {
+                multiplyIntCapped(TYPED_LETTER_MULTIPLIER, &matchWeight);
+            }
+            const bool isSameAsUserTypedLength = mInputLength == inputIndex + 1
+                    || (excessivePos == mInputLength - 1 && inputIndex == mInputLength - 2);
+            if (isSameAsUserTypedLength && isTerminal) {
+                const int freq = BinaryFormat::readFrequencyWithoutMovingPointer(DICT_ROOT, pos);
+                onTerminal(mWord, depth, DICT_ROOT, flags, pos, inputIndex, matchWeight, skipPos,
+                        excessivePos, transposedPos, freq, true, nextLetters, nextLettersSize);
+            }
+            // This character matched the typed character (enough to traverse the node at least)
+            // so we just evaluated it. Now we should evaluate this virtual node's children - that
+            // is, if it has any. If it has no children, we're done here - so we skip the end of
+            // the node, output the siblings position, and return false "don't traverse children".
+            // Note that !hasChildren implies isLastChar, so we know we don't have to skip any
+            // remaining char in this group for there can't be any.
+            if (!hasChildren) {
+                pos = BinaryFormat::skipFrequency(flags, pos);
+                *nextSiblingPosition =
+                        BinaryFormat::skipChildrenPosAndAttributes(DICT_ROOT, flags, pos);
+                return false;
+            }
+            // Start traversing all nodes after the index exceeds the user typed length
+            traverseAllNodes = isSameAsUserTypedLength;
+            diffs = diffs + ((NEAR_PROXIMITY_CHAR == matchedProximityCharId) ? 1 : 0);
+            // Finally, we are ready to go to the next character, the next "virtual node".
+            // We should advance the input index.
+            // We do this in this branch of the 'if traverseAllNodes' because we are still matching
+            // characters to input; the other branch is not matching them but searching for
+            // completions, this is why it does not have to do it.
+            ++inputIndex;
         }
-        if (!needsToTraverseChildrenNodes) return false;
-        *newTraverseAllNodes = traverseAllNodes;
-        *newMatchRate = matchWeight;
-        *newDiffs = diffs;
-        *newInputIndex = inputIndex;
-    } else {
-        const int *currentChars = getInputCharsAt(inputIndex);
-
-        if (transposedPos >= 0) {
-            if (inputIndex == transposedPos) currentChars += MAX_PROXIMITY_CHARS;
-            if (inputIndex == (transposedPos + 1)) currentChars -= MAX_PROXIMITY_CHARS;
+        // Optimization: Prune out words that are too long compared to how much was typed.
+        if (depth >= maxDepth || diffs > mMaxEditDistance) {
+            // We are giving up parsing this node and its children. Skip the rest of the node,
+            // output the sibling position, and return that we don't want to traverse children.
+            if (!isLastChar) {
+                pos = BinaryFormat::skipOtherCharacters(DICT_ROOT, pos);
+            }
+            pos = BinaryFormat::skipFrequency(flags, pos);
+            *nextSiblingPosition =
+                    BinaryFormat::skipChildrenPosAndAttributes(DICT_ROOT, flags, pos);
+            return false;
         }
 
-        int matchedProximityCharId = getMatchedProximityId(currentChars, c, skipPos, excessivePos,
-                transposedPos);
-        if (UNRELATED_CHAR == matchedProximityCharId) return false;
-        mWord[depth] = c;
-        // If inputIndex is greater than mInputLength, that means there is no
-        // proximity chars. So, we don't need to check proximity.
-        if (SAME_OR_ACCENTED_OR_CAPITALIZED_CHAR == matchedProximityCharId) {
-            multiplyIntCapped(TYPED_LETTER_MULTIPLIER, &matchWeight);
-        }
-        bool isSameAsUserTypedLength = mInputLength == inputIndex + 1
-                || (excessivePos == mInputLength - 1 && inputIndex == mInputLength - 2);
-        if (isSameAsUserTypedLength && terminal) {
-            onTerminal(mWord, depth, DICT_ROOT, flags, pos, inputIndex, matchWeight, skipPos,
-                    excessivePos, transposedPos, freq, true, nextLetters, nextLettersSize);
-        }
-        if (!needsToTraverseChildrenNodes) return false;
-        // Start traversing all nodes after the index exceeds the user typed length
-        *newTraverseAllNodes = isSameAsUserTypedLength;
-        *newMatchRate = matchWeight;
-        *newDiffs = diffs + ((NEAR_PROXIMITY_CHAR == matchedProximityCharId) ? 1 : 0);
-        *newInputIndex = inputIndex + 1;
-    }
-    // Optimization: Prune out words that are too long compared to how much was typed.
-    if (depth >= maxDepth || *newDiffs > mMaxEditDistance) {
-        return false;
-    }
+        // Prepare for the next character. Promote the prefetched char to current char - the loop
+        // will take care of prefetching the next. If we finally found our last char, nextc will
+        // contain NOT_A_CHARACTER.
+        c = nextc;
+        // Also, the next char is one "virtual node" depth more than this char.
+        ++depth;
+    } while (NOT_A_CHARACTER != c);
 
     // If inputIndex is greater than mInputLength, that means there are no proximity chars.
-    // TODO: Check if this can be isSameAsUserTypedLength only.
-    if (isSameAsUserTypedLength || mInputLength <= *newInputIndex) {
-        *newTraverseAllNodes = true;
+    // Here, that's all we are interested in so we don't need to check for isSameAsUserTypedLength.
+    if (mInputLength <= *newInputIndex) {
+        traverseAllNodes = true;
     }
-    // get the count of nodes and increment childAddress.
-    *newCount = Dictionary::getCount(DICT_ROOT, &childPosition);
-    *newChildPosition = childPosition;
-    if (DEBUG_DICT) assert(needsToTraverseChildrenNodes);
-    return needsToTraverseChildrenNodes;
+
+    // All the output values that are purely computation by this function are held in local
+    // variables. Output them to the caller.
+    *newTraverseAllNodes = traverseAllNodes;
+    *newMatchRate = matchWeight;
+    *newDiffs = diffs;
+    *newInputIndex = inputIndex;
+    *newOutputIndex = depth;
+
+    // Now we finished processing this node, and we want to traverse children. If there are no
+    // children, we can't come here.
+    assert(BinaryFormat::hasChildrenInFlags(flags));
+
+    // If this node was a terminal it still has the frequency under the pointer (it may have been
+    // read, but not skipped - see readFrequencyWithoutMovingPointer).
+    // Next come the children position, then possibly attributes (attributes are bigrams only for
+    // now, maybe something related to shortcuts in the future).
+    // Once this is read, we still need to output the number of nodes in the immediate children of
+    // this node, so we read and output it before returning true, as in "please traverse children".
+    pos = BinaryFormat::skipFrequency(flags, pos);
+    int childrenPos = BinaryFormat::readChildrenPosition(DICT_ROOT, flags, pos);
+    *nextSiblingPosition = BinaryFormat::skipChildrenPosAndAttributes(DICT_ROOT, flags, pos);
+    *newCount = BinaryFormat::getGroupCountAndForwardPointer(DICT_ROOT, &childrenPos);
+    *newChildrenPosition = childrenPos;
+    return true;
 }
 
 #endif // NEW_DICTIONARY_FORMAT
diff --git a/native/src/unigram_dictionary.h b/native/src/unigram_dictionary.h
index 789c495..dcc8f2a 100644
--- a/native/src/unigram_dictionary.h
+++ b/native/src/unigram_dictionary.h
@@ -36,10 +36,51 @@
     } ProximityType;
 
 public:
+#ifdef NEW_DICTIONARY_FORMAT
+
+    // Mask and flags for children address type selection.
+    static const int MASK_GROUP_ADDRESS_TYPE = 0xC0;
+    static const int FLAG_GROUP_ADDRESS_TYPE_NOADDRESS = 0x00;
+    static const int FLAG_GROUP_ADDRESS_TYPE_ONEBYTE = 0x40;
+    static const int FLAG_GROUP_ADDRESS_TYPE_TWOBYTES = 0x80;
+    static const int FLAG_GROUP_ADDRESS_TYPE_THREEBYTES = 0xC0;
+
+    // Flag for single/multiple char group
+    static const int FLAG_HAS_MULTIPLE_CHARS = 0x20;
+
+    // Flag for terminal groups
+    static const int FLAG_IS_TERMINAL = 0x10;
+
+    // Flag for bigram presence
+    static const int FLAG_HAS_BIGRAMS = 0x04;
+
+    // Attribute (bigram/shortcut) related flags:
+    // Flag for presence of more attributes
+    static const int FLAG_ATTRIBUTE_HAS_NEXT = 0x80;
+    // Flag for sign of offset. If this flag is set, the offset value must be negated.
+    static const int FLAG_ATTRIBUTE_OFFSET_NEGATIVE = 0x40;
+
+    // Mask for attribute frequency, stored on 4 bits inside the flags byte.
+    static const int MASK_ATTRIBUTE_FREQUENCY = 0x0F;
+
+    // Mask and flags for attribute address type selection.
+    static const int MASK_ATTRIBUTE_ADDRESS_TYPE = 0x30;
+    static const int FLAG_ATTRIBUTE_ADDRESS_TYPE_ONEBYTE = 0x10;
+    static const int FLAG_ATTRIBUTE_ADDRESS_TYPE_TWOBYTES = 0x20;
+    static const int FLAG_ATTRIBUTE_ADDRESS_TYPE_THREEBYTES = 0x30;
+#endif // NEW_DICTIONARY_FORMAT
+
     UnigramDictionary(const uint8_t* const streamStart, int typedLetterMultipler,
             int fullWordMultiplier, int maxWordLength, int maxWords, int maxProximityChars,
             const bool isLatestDictVersion);
+#ifndef NEW_DICTIONARY_FORMAT
     bool isValidWord(unsigned short *word, int length);
+#else // NEW_DICTIONARY_FORMAT
+    bool isValidWord(const uint16_t* const inWord, const int length) const;
+    int getBigrams(unsigned short *word, int length, int *codes, int codesSize,
+            unsigned short *outWords, int *frequencies, int maxWordLength, int maxBigrams,
+            int maxAlternatives);
+#endif // NEW_DICTIONARY_FORMAT
     int getBigramPosition(int pos, unsigned short *word, int offset, int length) const;
     int getSuggestions(const ProximityInfo *proximityInfo, const int *xcoordinates,
             const int *ycoordinates, const int *codes, const int codesSize, const int flags,
@@ -92,6 +133,7 @@
     }
     int getMostFrequentWordLike(const int startInputIndex, const int inputLength,
             unsigned short *word);
+#ifndef NEW_DICTIONARY_FORMAT
     void getWordsRec(const int childrenCount, const int pos, const int depth, const int maxDepth,
             const bool traverseAllNodes, const int snr, const int inputIndex, const int diffs,
             const int skipPos, const int excessivePos, const int transposedPos, int *nextLetters,
@@ -104,6 +146,11 @@
     bool processCurrentNodeForExactMatch(const int firstChildPos,
             const int startInputIndex, const int depth, unsigned short *word,
             int *newChildPosition, int *newCount, bool *newTerminal, int *newFreq, int *siblingPos);
+#else // NEW_DICTIONARY_FORMAT
+    int getFrequency(const uint16_t* const inWord, const int length) const;
+    int getMostFrequentWordLikeInner(const uint16_t* const inWord, const int length,
+            short unsigned int* outWord);
+#endif // NEW_DICTIONARY_FORMAT
 
     const uint8_t* const DICT_ROOT;
     const int MAX_WORD_LENGTH;