diff --git a/native/jni/NativeFileList.mk b/native/jni/NativeFileList.mk
index 18c6e7fd0..e4f6f0fe3 100755
--- a/native/jni/NativeFileList.mk
+++ b/native/jni/NativeFileList.mk
@@ -25,6 +25,7 @@ LATIN_IME_CORE_SRC_FILES := \
     ggml/common.cpp \
     ggml/context.cpp \
     ggml/gpt_neox.cpp \
+    ggml/LanguageModel.cpp \
     $(addprefix dictionary/header/, \
         header_policy.cpp \
         header_read_write_utils.cpp) \
diff --git a/native/jni/org_futo_inputmethod_latin_GGMLDictionary.cpp b/native/jni/org_futo_inputmethod_latin_GGMLDictionary.cpp
index 29612b8ac..c883a9c25 100644
--- a/native/jni/org_futo_inputmethod_latin_GGMLDictionary.cpp
+++ b/native/jni/org_futo_inputmethod_latin_GGMLDictionary.cpp
@@ -44,6 +44,9 @@
 #include "ggml/gpt_neox.h"
 #include "ggml/context.h"
 #include "ggml/common.h"
+#include "dictionary/structure/v2/patricia_trie_policy.h"
+#include "dictionary/structure/pt_common/dynamic_pt_reading_helper.h"
+#include "ggml/LanguageModel.h"
 
 #include <android/log.h>
 
@@ -150,9 +153,9 @@ float modifiedLevenshtein(const std::vector<KeyCoord>& a, const std::vector<KeyC
 
 
 // TODO: https://www.npmjs.com/package/fastest-levenshtein?activeTab=code
-int levenshtein(const std::string &a, const std::string &b) {
-    int a_len = a.length();
-    int b_len = b.length();
+int levenshtein(const char *a, const char *b, size_t len) {
+    size_t a_len = len;
+    size_t b_len = len;
 
     // Initialize matrix of zeros
     std::vector<std::vector<int>> d(a_len + 1, std::vector<int>(b_len + 1, 0));
@@ -198,50 +201,66 @@ static std::string trim(const std::string &s) {
 namespace latinime {
 
 struct DictionaryRescorer {
-    std::vector<std::vector<std::string>> id_to_word;
+    bool initialized = false;
+
+    // TODO: We should store dictionary here too to look up words during multi-token sampling
+
+    std::vector<int> tokenStrategies;
+
+    std::unordered_set<int> invalidTokens;
+
+    // TODO: words like "won't", "can't", are tokenized like won, 't, can, 't or won, ', t, can, ', t
+    //       By taking the first token and assuming it's done we get nonsensical suggestions like won, can,
+    std::unordered_set<int> wordTokens;
+
+
+    std::unordered_set<int> continueSamplingTokens;
+    std::unordered_set<int> continuationToken;
+
 };
 
-void DictionaryRescorer_addDictionary(Dictionary &dict, gpt_vocab &vocab, DictionaryRescorer &rescorer) {
-    if(rescorer.id_to_word.size() < vocab.id_to_token.size()) {
-        rescorer.id_to_word.resize(vocab.id_to_token.size());
+#define STRATEGY_CONTINUATION 4
+void DictionaryRescorer_addDictionary(Dictionary &dict, const LanguageModel &model, DictionaryRescorer &rescorer) {
+    if(rescorer.initialized) return;
+
+    rescorer.tokenStrategies.clear();
+
+    if(rescorer.tokenStrategies.size() < model.getVocabSize()) {
+        rescorer.tokenStrategies.resize(model.getVocabSize());
     }
-    int token = 0;
 
-    int wordCodePoints[MAX_WORD_LENGTH];
-    int wordCodePointCount = 0;
+    for(int i=0; i<model.getVocabSize(); i++) {
+        const char *word = model.getToken(i);
 
-    char word_c[MAX_WORD_LENGTH * 4];
+        char c = word[0];
 
-    AKLOGI("Adding words..");
-    int n = 0;
-    do {
-        n++;
-        token = dict.getNextWordAndNextToken(token, wordCodePoints, &wordCodePointCount);
+        bool startOfWord = c == ' ';
 
-        bool isBeginningOfSentence = false;
-        if (wordCodePointCount > 0 && wordCodePoints[0] == CODE_POINT_BEGINNING_OF_SENTENCE) {
-            isBeginningOfSentence = true;
+        bool isInvalid = c == ',' || c == '.' || c == '?' || c == '!' || ((c >= '0') && (c <= '9')) || c == '(' || c == ')' || c == '"' || c == '[' || c == ']' || c == '+' || c == '#' || c == '<' || c == '>' || c == '|';
+
+        // TODO: The dictionary never contains numbers
+        const int strategy = isInvalid ? STRATEGY_INVALID : (startOfWord ? dict.getWordStrategy(word + 1) : STRATEGY_CONTINUATION);
+        rescorer.tokenStrategies[i] = strategy;
+        switch(strategy) {
+            case STRATEGY_COMMIT_WORD:
+                rescorer.wordTokens.insert(i);
+                break;
+            case STRATEGY_CONTINUE_SAMPLING:
+                // TODO: We may need something like
+                rescorer.continueSamplingTokens.insert(i);
+                break;
+            case STRATEGY_INVALID:
+                rescorer.invalidTokens.insert(i);
+                break;
+            case STRATEGY_CONTINUATION:
+                rescorer.continuationToken.insert(i);
+                break;
+            default:
+                ASSERT(false);
         }
+    }
 
-        intArrayToCharArray(
-                isBeginningOfSentence ? wordCodePoints + 1 : wordCodePoints,
-                isBeginningOfSentence ? wordCodePointCount - 1 : wordCodePointCount,
-                word_c,
-                MAX_WORD_LENGTH * 4
-        );
-
-        std::string word(word_c);
-
-        word = std::string(" ") + trim(word);
-
-
-        std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, word);
-        gpt_vocab::id key = tokens[0];
-
-        rescorer.id_to_word[key].push_back(word);
-    } while(token != 0);
-
-    AKLOGI("Added %d words\n", n);
+    rescorer.initialized = true;
 }
 
 template<typename T>
@@ -251,19 +270,143 @@ bool sortProbabilityPairDescending(const std::pair<float, T>& a, const std::pair
 
 
 template<typename T>
-static inline void sortProbabilityPairVectorDescending(std::vector<std::pair<float, T>> vec) {
+static inline void sortProbabilityPairVectorDescending(std::vector<std::pair<float, T>> &vec) {
     std::sort(vec.begin(), vec.end(), sortProbabilityPairDescending<T>);
 }
 
+template<typename T>
+static inline void sortProbabilityPairVectorDescending(std::vector<std::pair<float, T>> &vec, int partial) {
+    std::partial_sort(vec.begin(), vec.begin() + partial, vec.end(), sortProbabilityPairDescending<T>);
+}
+
+
+
+float rescore_token_levenshtein(float prob, const char *text, int length, const std::string &partialWord, bool applyLengthPenalty) {
+    if(prob == 0.0f) return 0.0f;
+
+    if(!partialWord.empty()) {
+        unsigned int min_length = std::min(length, partialWord.length());
+        float distance = (float)levenshtein(text, partialWord.c_str(), min_length);
+
+        if(applyLengthPenalty && (length < partialWord.length())) {
+            distance += (partialWord.length() - length) * 2.0f;
+        }
+
+        // this assumes the probabilities are all positive
+        ASSERT(prob >= 0.0f;)
+
+        prob = prob / (1.0f + distance);
+
+        return prob;
+    } else {
+        return prob;
+    }
+}
+
+float rescore_token_levenshtein(float prob, const std::string &text, const std::string &partialWord, bool applyLengthPenalty) {
+    return rescore_token_levenshtein(prob, text.c_str(), text.length(), partialWord, applyLengthPenalty);
+}
+
+
+std::pair<float, token_sequence> process_token_sequence(    
+    const DictionaryRescorer &rescorer,
+    LanguageModel &model,
+    const std::string &partialWord,
+    const token_sequence &seq,
+    float lastprob,
+    float minprob,
+    int recursionDepth
+) {
+    if(recursionDepth > 3) {
+        // Cut our losses and exit
+        return { 0.0f, {} };
+    }
+
+    std::vector<float> nextLogits = model.temporarilyInfer(seq);
+    std::vector<std::pair<float, int>> nextIndexValue;
+
+    for (int j = 0; j < nextLogits.size(); j++) {
+        int thisStrategy = rescorer.tokenStrategies[j];
+
+        nextIndexValue.emplace_back(nextLogits[j], j);
+    }
+    
+    sortProbabilityPairVectorDescending(nextIndexValue, 3);
+    for (int j = 0; j < 3; j++) {
+        float probability = nextIndexValue[j].first;
+        int tokenId = nextIndexValue[j].second;
+
+        const char * chars = model.getToken(tokenId);
+
+        // handle punctuation and stuf as well
+        // we really need an abstract function to return token type
+        if(chars[0] == ' ') {
+            // The model believes the previous word has ended, so let's just cut it here and return verbatim
+            // TODO: should lastprob be modified with the probability value? what if we reach this only in the
+            //       3rd iteration?
+            return { lastprob, seq };
+        }
+
+        token_sequence new_token_sequence = seq + { nextToken };
+
+        std::string resultingWord = model.decode(new_token_sequence);
+
+        // Rescore according to partial word, if exists
+        probability = rescore_token_levenshtein(probability, resultingWord, partialWord, true);
+
+        // We do this AFTER rescoring as lastprob is also after rescoring
+        // TODO: Is a simple average sufficient here? What about during recursion?
+        float resultingProbability = (probability + lastprob) / 2.0f;
+
+        if(resultingProbability < minprob) continue;
+
+        // TODO: Check with the dictionary here. Need to remember pt position for optimization
+        // (pass pt from the function)
+        int strategy = TODO(check strategy for this now)
+
+        if(strategy == STRATEGY_COMMIT_WORD) {
+            // We've finally written a word, so we can return this
+            return { resultingProbability, new_token_sequence }
+        }else if(strategy == STRATEGY_CONTINUE_SAMPLING) {
+            return process_token_sequence(rescorer, model, partialWord, new_token_sequence, resultingProbability, minprob, recursionDepth+1);
+        }else{
+            // The dictionary says this is an invalid word. We can do two things here
+            // 1. Trust the dictionary - Discard it, because we will never arrive at a word in the dictionary
+            // 2. Trust the model - Continue sampling until space, we trust that the model is giving something useful
+            //    (a special word like JQuery that may not be in the dictionary)
+            // A problem for #2 is that we ignore invalid words anyway when it's the first token
+        }
+    }
+
+    return { 0.0f, {} };
+}
+
 std::vector<std::pair<float, std::string>> DictionaryRescorer_process(
         const DictionaryRescorer &rescorer,
-        const std::vector<float> &logits,
+        const std::vector<float> &logitsOrig,
+        const std::unordered_set<gpt_vocab::id> &punctIds,
         const std::string &partialWord,
-        gpt_vocab &vocab,
+        LanguageModel &model,
         int n
 ) {
     std::vector<std::pair<float, std::string>> top_n_results(n);
 
+    std::vector<float> logits(logitsOrig);
+
+    for(int i : rescorer.invalidTokens) {
+        logits[i] = 0.0f;
+    }
+
+    for(int i : rescorer.continuationToken) {
+        logits[i] = 0.0f; // TODO: Allow continuation only if it's the most probable token, and if partialWord is empty
+    }
+
+    // Restore punctuation
+    // TODO: A better way
+    for(int i : punctIds){
+        logits[i] = logitsOrig[i];
+    }
+
     // Get a vector of index and value pairs
     std::vector<std::pair<float, int>> index_value;
     for (int i = 0; i < logits.size(); i++) {
@@ -271,54 +414,59 @@ std::vector<std::pair<float, std::string>> DictionaryRescorer_process(
     }
 
     // Sort the index_value vector in descending order of value
-    sortProbabilityPairVectorDescending(index_value);
+    sortProbabilityPairVectorDescending(index_value, 6000);
 
     if(!partialWord.empty()) {
-        // TODO: Figure out a better way
-        index_value.resize(1000);
+        // TODO: Figure out a better way. It's slow to compute full levenshtein for every value, and inaccurate to resize to a small size like 500.
+        // We could cache past distances to prune those that are too distant
+        // We could compare first characters of words and pick those as more likely
+        index_value.resize(std::min((size_t)6000, rescorer.wordTokens.size()));
         // Adjust probabilities according to levenshtein distance
         for(auto &v : index_value) {
             int token_id = v.second;
+            int thisStrategy = rescorer.tokenStrategies[token_id];
 
-            // String based
-            std::string token = vocab.id_to_token[token_id];
+            const char *token = model.getToken(token_id);
+            size_t token_length = strlen(token);
 
-            unsigned int min_length = std::min(token.length(), partialWord.length());
-
-            float distance = (float)levenshtein(token.substr(0, min_length), partialWord.substr(0, min_length));
-
-            // this assumes the probabilities are all positive
-            v.first = v.first / (1.0f + distance);
+            // Apply length penalty for when the token is too short, except when its continue_sampling (why?)
+            v.first = rescore_token_levenshtein(v.first, token, token_length, partialWord, (thisStrategy != STRATEGY_CONTINUE_SAMPLING || (token_length < 3)) );
         }
 
         // Sort the index_value vector in descending order of value again
-        sortProbabilityPairVectorDescending(index_value);
+        sortProbabilityPairVectorDescending(index_value, n);
     }
 
-    index_value.resize(100);
+    std::vector<std::pair<float, token_sequence>> top_three_results_so_far(3);
 
-    for(auto & v : index_value){
-        gpt_vocab::id token_id = v.second;
+    // Select the top three results we can commit instantly
+    for(int i=0; i<n; i++) {
+        float probability = index_value[i].first;
+        int tokenId = index_value[i].second;
 
-        for(const std::string& str : rescorer.id_to_word[token_id]) {
-            top_n_results.emplace_back(v.first, str);
+        int strategy = rescorer.tokenStrategies[tokenId];
+
+        if(strategy == STRATEGY_COMMIT_WORD) {
+            top_three_results_so_far.emplace_back(probability, { tokenId });
+            if(top_three_results_so_far.size() >= 3) break;
         }
     }
 
+    // Iterate over those that require continuing sampling (top three only (TODO?))
+    for(int i=0; i<std::min(3, n); i++) {
+        float probability = index_value[i].first;
+        int tokenId = index_value[i].second;
 
-    if(!partialWord.empty()) {
-        // Adjust probabilities according to levenshtein distance
-        for(auto &v : top_n_results) {
-            unsigned int min_length = std::min(v.second.length(), partialWord.length());
+        int strategy = rescorer.tokenStrategies[tokenId];
 
-            float distance = (float)levenshtein(v.second.substr(0, min_length), partialWord.substr(0, min_length));
+        if(strategy != STRATEGY_CONTINUE_SAMPLING) continue;
+        if(!top_three_results_so_far.empty() && probability < top_three_results_so_far.back().first) continue;
 
-            // this assumes the probabilities are all positive
-            v.first = v.first / (1.0f + distance);
+        auto result = process_token_sequence(rescorer, model, partialWord, { tokenId }, probability, top_three_results_so_far.back().first, 0);
+        if(result.first != 0.0f) {
+            top_three_results_so_far.push_back(result);
+            sortProbabilityPairVectorDescending(top_three_results_so_far);
         }
-
-        // Sort the top_n_vector vector in descending order of probability
-        sortProbabilityPairVectorDescending(top_n_results);
     }
 
     return top_n_results;
@@ -327,21 +475,13 @@ std::vector<std::pair<float, std::string>> DictionaryRescorer_process(
 
 
 struct GGMLDictionaryState {
-    int n_threads = 3;
+    LanguageModel *model;
 
-    transformer_context t_context;
-
-    std::vector<float> logits;
-    std::vector<gpt_vocab::id> bad_logits;
-    std::unordered_set<gpt_vocab::id> punct_logits;
+    std::vector<gpt_vocab::id> bad_ids;
+    std::unordered_set<gpt_vocab::id> punct_ids;
 
     //std::map<ProximityInfo *, KeyboardVocab> proximity_info_to_kvoc;
     DictionaryRescorer rescorer;
-
-    size_t mem_per_token = 0;
-
-    gpt_neox_model model;
-    gpt_vocab vocab;
 };
 
 static jlong latinime_GGMLDictionary_open(JNIEnv *env, jclass clazz, jstring sourceDir,
@@ -359,18 +499,15 @@ static jlong latinime_GGMLDictionary_open(JNIEnv *env, jclass clazz, jstring sou
 
     GGMLDictionaryState *state = new GGMLDictionaryState();
 
-    std::string fname(sourceDirChars);
-
-    bool result = gpt_neox_model_load(fname, state->model, state->vocab);
-
-    if(!result) {
+    state->model = GPTNeoXAdapter::createLanguageModel(sourceDirChars);
+    if(!state->model) {
         AKLOGE("GGMLDict: Could not load model");
         free(state);
         return 0;
     }
 
-    for(int i=0; i<state->model.hparams.n_vocab; i++){
-        std::string token = state->vocab.id_to_token[i];
+    for(int i=0; i<state->model->getVocabSize(); i++){
+        std::string token = state->model->getToken(i);
 
         bool is_bad = token.empty();
         bool has_punct = false;
@@ -381,7 +518,7 @@ static jlong latinime_GGMLDictionary_open(JNIEnv *env, jclass clazz, jstring sou
                 bool is_punct = c == ',' || c == '.' || c == '?' || c == '!';
                 bool is_letter = ((c >= 'a') && (c <= 'z')) || ((c >= 'A') && (c <= 'Z'));
                 bool is_number = (c >= '0') && (c <= '9');
-                bool is_special = c == '(' || c == ')' || c == '"' || c == '[' || c == ']' || c == '+' || c == '#';
+                bool is_special = c == '(' || c == ')' || c == '"' || c == '[' || c == ']' || c == '+' || c == '#' || c == '<' || c == '>';
 
                 if(is_punct || is_special) has_punct = true;
 
@@ -398,15 +535,13 @@ static jlong latinime_GGMLDictionary_open(JNIEnv *env, jclass clazz, jstring sou
         is_bad = is_bad || num_chars == 0;
 
         if(is_bad) {
-            state->bad_logits.emplace_back(i);
+            state->bad_ids.emplace_back(i);
         }
         if(has_punct) {
-            state->punct_logits.insert(i);
+            state->punct_ids.insert(i);
         }
     }
 
-
-
     PROF_TIMER_END(66);
     return reinterpret_cast<jlong>(state);
 }
@@ -419,16 +554,19 @@ static void latinime_GGMLDictionary_close(JNIEnv *env, jclass clazz, jlong dict)
 
 
 static void latinime_GGMLDictionary_addDict(JNIEnv *env, jclass clazz, jlong statePtr, jlong dict) {
-    AKLOGI("Adding dictionary %ld\n", dict);
     GGMLDictionaryState *state = reinterpret_cast<GGMLDictionaryState *>(statePtr);
     Dictionary *dictionary = reinterpret_cast<Dictionary *>(dict);
 
-    AKLOGI("Here is the dictionary we ading:");
+    AKLOGI("Here is the dictionary we are adding:");
     dictionary->logDictionaryInfo(env);
 
-    DictionaryRescorer_addDictionary(*dictionary, state->vocab, state->rescorer);
+    time_t t1 = time(NULL);
+    DictionaryRescorer_addDictionary(*dictionary, *state->model, state->rescorer);
+    time_t t2 = time(NULL);
+    AKLOGI("Took %.2f to add dictionary", difftime(t2, t1));
 }
 
+
 static void latinime_GGMLDictionary_getSuggestions(JNIEnv *env, jclass clazz,
        // inputs
        jlong dict,
@@ -470,59 +608,37 @@ static void latinime_GGMLDictionary_getSuggestions(JNIEnv *env, jclass clazz,
         env->ReleaseStringUTFChars(partialWord, pwstr);
     }
 
-    token_sequence next_context = gpt_tokenize(state->vocab, contextString);
+    token_sequence next_context = state->model->tokenize(contextString);
+    bool allow_punctuation_next = state->punct_ids.count(next_context[next_context.size() - 1]) == 0;
 
-    bool allow_punctuation_next = state->punct_logits.count(next_context[next_context.size() - 1]) == 0;
+    state->model->updateContext(next_context);
+    std::vector<float> logits = state->model->infer();
 
-    //truncate to front of the prompt if its too long
-    int32_t nctx = state->model.hparams.n_ctx;
-
-    if (next_context.size() + 2 > nctx) {
-        int offset = next_context.size() - nctx + 2;
-        next_context = std::vector<int>(next_context.begin() + offset, next_context.end());
-    }
-
-
-    auto fastforward_info = transformer_context_fastforward(state->t_context, next_context);
-
-    token_sequence &embd_inp = fastforward_info.first;
-    int n_past = fastforward_info.second;
-
-    if(!embd_inp.empty()) {
-        AKLOGI("npast = %d, size(embd) = %d\n", n_past, (int) embd_inp.size());
-        gpt_neox_eval(state->model, state->n_threads, n_past, embd_inp, state->logits,
-                      state->mem_per_token);
-
-        transformer_context_apply(state->t_context, fastforward_info);
-    }
-
-    int topid = std::min_element(state->logits.begin(),state->logits.end())-state->logits.begin();
-    float zeroValue = (state->logits[topid] < 0 ? state->logits[topid] : 0);
-
-    for(int bad_id : state->bad_logits) {
-        state->logits[bad_id] = zeroValue;
+    float zeroValue = 0.0f;
+    for(int bad_id : state->bad_ids) {
+        logits[bad_id] = zeroValue;
     }
 
     // Don't allow punctuation after we just wrote punctuation
     if(!allow_punctuation_next) {
-        for(int bad_id : state->punct_logits) {
-            state->logits[bad_id] = zeroValue;
+        for(int bad_id : state->punct_ids) {
+            logits[bad_id] = zeroValue;
         }
     }
 
-    auto results = DictionaryRescorer_process(state->rescorer, state->logits, partialWordString, state->vocab, 10);
-
+    auto results = DictionaryRescorer_process(state->rescorer, logits, state->punct_ids, partialWordString, *state->model, 50);
 
     size_t size = env->GetArrayLength(outPredictions);
 
     // Get the array elements
     jfloat *probsArray = env->GetFloatArrayElements(outProbabilities, nullptr);
 
+    AKLOGI("Predictions:");
     // Output predictions for next word
     for (int i = 0; i < std::min(size, results.size()); i++) {
         std::string &word = results[i].second;
         if (i < 8) {
-            AKLOGI(" - prediction[%d]: %s", i, word.c_str());
+            AKLOGI(" - prediction[%d]: [%s]", i, word.c_str());
         }
         jstring jstr = env->NewStringUTF(word.c_str());
 
diff --git a/native/jni/src/dictionary/interface/dictionary_structure_with_buffer_policy.h b/native/jni/src/dictionary/interface/dictionary_structure_with_buffer_policy.h
index ace48491d..8a9a15f83 100644
--- a/native/jni/src/dictionary/interface/dictionary_structure_with_buffer_policy.h
+++ b/native/jni/src/dictionary/interface/dictionary_structure_with_buffer_policy.h
@@ -114,6 +114,8 @@ class DictionaryStructureWithBufferPolicy {
 
     virtual bool isCorrupted() const = 0;
 
+    virtual int getWordStrategy(const char *text) const = 0;
+
  protected:
     DictionaryStructureWithBufferPolicy() {}
 
diff --git a/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.cpp b/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.cpp
index 6fb9cffb7..a86f79ee4 100644
--- a/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.cpp
+++ b/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.cpp
@@ -657,6 +657,18 @@ int Ver4PatriciaTriePolicy::getTerminalPtNodePosFromWordId(const int wordId) con
     return wordId == NOT_A_WORD_ID ? NOT_A_DICT_POS : wordId;
 }
 
+int Ver4PatriciaTriePolicy::getWordStrategy(const char *text) const {
+    DynamicPtReadingHelper readingHelper(&mNodeReader, &mPtNodeArrayReader);
+    readingHelper.initWithPtNodeArrayPos(getRootPosition());
+    const int strategy = readingHelper.searchWordAndReturnStrategy(text);
+    if (readingHelper.isError()) {
+        mIsCorrupted = true;
+        AKLOGE("Dictionary reading error in getWordId().");
+    }
+    return strategy;
+}
+
+
 } // namespace v402
 } // namespace backward
 } // namespace latinime
diff --git a/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.h b/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.h
index bce5f6bea..e284eb192 100644
--- a/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.h
+++ b/native/jni/src/dictionary/structure/backward/v402/ver4_patricia_trie_policy.h
@@ -139,6 +139,7 @@ class Ver4PatriciaTriePolicy : public DictionaryStructureWithBufferPolicy {
         return mIsCorrupted;
     }
 
+    int getWordStrategy(const char *text) const;
  private:
     DISALLOW_IMPLICIT_CONSTRUCTORS(Ver4PatriciaTriePolicy);
 
diff --git a/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.cpp b/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
index 6f96a5a0b..4d3629d81 100644
--- a/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
+++ b/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.cpp
@@ -600,4 +600,15 @@ int Ver4PatriciaTriePolicy::getNextWordAndNextToken(const int token, int *const
     return nextToken;
 }
 
+int Ver4PatriciaTriePolicy::getWordStrategy(const char *text) const {
+    DynamicPtReadingHelper readingHelper(&mNodeReader, &mPtNodeArrayReader);
+    readingHelper.initWithPtNodeArrayPos(getRootPosition());
+    const int strategy = readingHelper.searchWordAndReturnStrategy(text);
+    if (readingHelper.isError()) {
+        mIsCorrupted = true;
+        AKLOGE("Dictionary reading error in createAndGetAllChildDicNodes().");
+    }
+    return strategy;
+}
+
 } // namespace latinime
diff --git a/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.h b/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.h
index d130a4e78..79eba3978 100644
--- a/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.h
+++ b/native/jni/src/dictionary/structure/v4/ver4_patricia_trie_policy.h
@@ -118,6 +118,7 @@ class Ver4PatriciaTriePolicy : public DictionaryStructureWithBufferPolicy {
         return mIsCorrupted;
     }
 
+    int getWordStrategy(const char *text) const;
  private:
     DISALLOW_IMPLICIT_CONSTRUCTORS(Ver4PatriciaTriePolicy);
 
diff --git a/native/jni/src/ggml/LanguageModel.cpp b/native/jni/src/ggml/LanguageModel.cpp
new file mode 100644
index 000000000..6bda423e3
--- /dev/null
+++ b/native/jni/src/ggml/LanguageModel.cpp
@@ -0,0 +1,55 @@
+//
+// Created by alex on 7/24/23.
+//
+
+#include "LanguageModel.h"
+
+LanguageModelAdapter::~LanguageModelAdapter() {};
+
+LanguageModel::LanguageModel(LanguageModelAdapter *adapter): adapter(adapter) {
+
+}
+
+int GPTNeoXAdapter::getVocabSize() const {
+    return model.hparams.n_vocab;
+}
+
+const char *GPTNeoXAdapter::getToken(int id) const {
+    return vocab.id_to_token.at(id).c_str();
+}
+
+bool GPTNeoXAdapter::eval(int nPast, token_sequence input, std::vector<float> &outLogits) {
+    // TODO
+    ASSERT(nPast + input.size() < model.hparams.n_ctx);
+
+    return gpt_neox_eval(model, numThreads, nPast, input, outLogits, memPerToken);
+}
+
+std::vector<int> GPTNeoXAdapter::tokenize(const char *text) {
+    return gpt_tokenize(vocab, text);
+}
+
+std::string GPTNeoXAdapter::decode(const token_sequence &tokens) const {
+    // For now we just merge the tokens together, this may need to be different for other languages and unicode
+    size_t length = 0;
+    for(int token : tokens) length += strlen(getToken(token));
+
+    std::string result(length);
+    for(int token : tokens) result.append(getToken(token));
+
+    return result;
+}
+
+LanguageModel *GPTNeoXAdapter::createLanguageModel(const char *path) {
+    auto adapter = new GPTNeoXAdapter();
+
+    bool result = gpt_neox_model_load(path, adapter->model, adapter->vocab);
+    if(!result) {
+        delete adapter;
+        return nullptr;
+    }
+
+    return new LanguageModel(adapter);
+}
+
+GPTNeoXAdapter::GPTNeoXAdapter() = default;
diff --git a/native/jni/src/ggml/LanguageModel.h b/native/jni/src/ggml/LanguageModel.h
new file mode 100644
index 000000000..cfda57808
--- /dev/null
+++ b/native/jni/src/ggml/LanguageModel.h
@@ -0,0 +1,136 @@
+//
+// Created by alex on 7/24/23.
+//
+
+#ifndef LATINIME_LANGUAGEMODEL_H
+#define LATINIME_LANGUAGEMODEL_H
+
+#include <vector>
+#include <unordered_set>
+#include "context.h"
+#include "defines.h"
+#include "gpt_neox.h"
+
+class LanguageModelAdapter {
+public:
+    int numThreads = 4;
+
+    virtual int getVocabSize() const = 0;
+    virtual const char *getToken(int id) const = 0;
+    virtual bool eval(int nPast, token_sequence input, std::vector<float> &outLogits) = 0;
+
+    virtual std::vector<int> tokenize(const char *text) = 0;
+    virtual std::string decode(const token_sequence &tokens) const = 0;
+
+    virtual ~LanguageModelAdapter() = 0;
+};
+
+class LanguageModel {
+public:
+    LanguageModel(LanguageModelAdapter *adapter);
+
+    // Tokenizes the given text to tokens
+    AK_FORCE_INLINE std::vector<int> tokenize(const char *text) const {
+        return adapter->tokenize(text);
+    }
+    AK_FORCE_INLINE std::vector<int> tokenize(const std::string &text) const {
+        return tokenize(text.c_str());
+    }
+
+    AK_FORCE_INLINE std::string decode(const token_sequence &tokens) const {
+        return adapter->decode(tokens);
+    }
+
+    // Fast forward the context
+    AK_FORCE_INLINE void updateContext(const std::vector<int> &newContext) {
+        auto result = transformer_context_fastforward(transformerContext, newContext);
+        pendingEvaluationSequence = result.first;
+        pendingNPast = result.second;
+
+        pendingContext = newContext;
+    }
+    AK_FORCE_INLINE void updateContext(const char *text) {
+        return updateContext(tokenize(text));
+    }
+
+    AK_FORCE_INLINE void pushToContext(int token) {
+        pendingContext.push_back(token);
+        updateContext(pendingContext);
+    }
+
+    // TODO: This method returns a copy of 128kB of data
+    AK_FORCE_INLINE std::vector<float> infer() {
+        if(pendingEvaluationSequence.empty()){
+            AKLOGI("LanguageModel: evaluation skipped due to empty pending evaluation sequence\n");
+            return outLogits;
+        }
+
+        if(!adapter->eval(pendingNPast, pendingEvaluationSequence, outLogits)) {
+            ASSERT(false);
+        }
+
+        transformer_context_apply(transformerContext, {pendingEvaluationSequence, pendingNPast});
+
+        pendingEvaluationSequence.clear();
+
+        return outLogits;
+    }
+
+    // Infers the given tokens on top of the active context without updating cache.
+    // TODO: This method returns a copy of 128kB of data
+    AK_FORCE_INLINE std::vector<float> temporarilyInfer(const std::vector<int> &tokens) {
+        ASSERT(pendingEvaluationSequence.empty());
+        ASSERT(!tokens.empty());
+
+        if(!adapter->eval(transformerContext.active_context.size(), tokens, tmpOutLogits)) {
+            ASSERT(false);
+        }
+
+        return tmpOutLogits;
+    }
+
+    AK_FORCE_INLINE int getVocabSize() const {
+        return adapter->getVocabSize();
+    }
+
+    AK_FORCE_INLINE const char *getToken(int token) const {
+        return adapter->getToken(token);
+    }
+
+    AK_FORCE_INLINE bool isPendingEvaluation() const {
+        return pendingEvaluationSequence.size() > 0;
+    }
+private:
+    token_sequence pendingContext;
+    token_sequence pendingEvaluationSequence;
+    int pendingNPast = 0;
+
+    LanguageModelAdapter *adapter;
+
+    transformer_context transformerContext;
+
+    std::vector<float> outLogits;
+    std::vector<float> tmpOutLogits;
+
+    std::unordered_set<int> punctIds;
+};
+
+
+class GPTNeoXAdapter : public LanguageModelAdapter {
+public:
+    int getVocabSize() const;
+    const char *getToken(int id) const;
+    bool eval(int nPast, token_sequence input, std::vector<float> &outLogits);
+    virtual std::vector<int> tokenize(const char *text);
+    virtual std::string decode(const token_sequence &tokens) const;
+
+    static LanguageModel *createLanguageModel(const char *path);
+private:
+    GPTNeoXAdapter();
+    gpt_vocab vocab;
+    gpt_neox_model model;
+
+    size_t memPerToken = 0;
+};
+
+#endif //LATINIME_LANGUAGEMODEL_H
diff --git a/native/jni/src/suggest/core/dictionary/dictionary.cpp b/native/jni/src/suggest/core/dictionary/dictionary.cpp
index 5c9a1392e..11356bc7e 100644
--- a/native/jni/src/suggest/core/dictionary/dictionary.cpp
+++ b/native/jni/src/suggest/core/dictionary/dictionary.cpp
@@ -196,6 +196,11 @@ int Dictionary::getNextWordAndNextToken(const int token, int *const outCodePoint
             token, outCodePoints, outCodePointCount);
 }
 
+int Dictionary::getWordStrategy(const char *text) const {
+    TimeKeeper::setCurrentTime();
+    return mDictionaryStructureWithBufferPolicy->getWordStrategy(text);
+}
+
 void Dictionary::logDictionaryInfo(JNIEnv *const env) const {
     int dictionaryIdCodePointBuffer[HEADER_ATTRIBUTE_BUFFER_SIZE];
     int versionStringCodePointBuffer[HEADER_ATTRIBUTE_BUFFER_SIZE];
diff --git a/native/jni/src/suggest/core/dictionary/dictionary.h b/native/jni/src/suggest/core/dictionary/dictionary.h
index a01a705d2..0ed60f030 100644
--- a/native/jni/src/suggest/core/dictionary/dictionary.h
+++ b/native/jni/src/suggest/core/dictionary/dictionary.h
@@ -118,6 +118,7 @@ class Dictionary {
 
     void logDictionaryInfo(JNIEnv *const env) const;
 
+    int getWordStrategy(const char *word) const;
 private:
     DISALLOW_IMPLICIT_CONSTRUCTORS(Dictionary);