diff --git a/native_client/BUILD b/native_client/BUILD
index a64ff2be..5d001c96 100644
--- a/native_client/BUILD
+++ b/native_client/BUILD
@@ -67,6 +67,7 @@ tf_cc_shared_object(
         ### Obtained by trial/error process ...
         ### CPU only build libdeepspeech.so from 63M to 36M
         "//tensorflow/core/kernels:constant_op",        # Const
+        "//tensorflow/core/kernels:immutable_constant_op", # ImmutableConst
         "//tensorflow/core/kernels:identity_op",        # Identity
         "//tensorflow/core/kernels:transpose_op",       # Transpose
         "//tensorflow/core/kernels:reshape_op",         # Reshape
diff --git a/native_client/deepspeech.cc b/native_client/deepspeech.cc
index 02c18ecf..3450572d 100644
--- a/native_client/deepspeech.cc
+++ b/native_client/deepspeech.cc
@@ -16,6 +16,7 @@
 
 #include "tensorflow/core/public/session.h"
 #include "tensorflow/core/platform/env.h"
+#include "tensorflow/core/util/memmapped_file_system.h"
 
 #define BATCH_SIZE 1
 
@@ -27,6 +28,7 @@ namespace DeepSpeech {
 
 class Private {
   public:
+    MemmappedEnv* mmap_env;
     Session* session;
     GraphDef graph_def;
     int ncep;
@@ -42,6 +44,7 @@ Model::Model(const char* aModelPath, int aNCep, int aNContext,
              const char* aAlphabetConfigPath, int aBeamWidth)
 {
   mPriv             = new Private;
+  mPriv->mmap_env   = new MemmappedEnv(Env::Default());
   mPriv->session    = NULL;
   mPriv->scorer     = NULL;
   mPriv->ncep       = aNCep;
@@ -56,13 +59,39 @@ Model::Model(const char* aModelPath, int aNCep, int aNContext,
     return;
   }
 
-  Status status = NewSession(SessionOptions(), &mPriv->session);
+  Status status;
+  SessionOptions options;
+  bool is_mmap = std::string(aModelPath).find(".pbmm") != std::string::npos;
+  if (!is_mmap) {
+    std::cerr << "Warning: reading entire model file into memory. Transform model file into an mmapped graph to reduce heap usage." << std::endl;
+  }
+
+  if (is_mmap) {
+    status = mPriv->mmap_env->InitializeFromFile(aModelPath);
+    if (!status.ok()) {
+      std::cerr << status.ToString() << std::endl;
+      return;
+    }
+
+    options.config.mutable_graph_options()
+      ->mutable_optimizer_options()
+      ->set_opt_level(::OptimizerOptions::L0);
+    options.env = mPriv->mmap_env;
+  }
+
+  status = NewSession(options, &mPriv->session);
   if (!status.ok()) {
     std::cerr << status.ToString() << std::endl;
     return;
   }
 
-  status = ReadBinaryProto(Env::Default(), aModelPath, &mPriv->graph_def);
+  if (is_mmap) {
+    status = ReadBinaryProto(mPriv->mmap_env,
+                             MemmappedFileSystem::kMemmappedPackageDefaultGraphDef,
+                             &mPriv->graph_def);
+  } else {
+    status = ReadBinaryProto(Env::Default(), aModelPath, &mPriv->graph_def);
+  }
   if (!status.ok()) {
     mPriv->session->Close();
     mPriv->session = NULL;
@@ -105,6 +134,7 @@ Model::~Model()
     mPriv->session->Close();
   }
 
+  delete mPriv->mmap_env;
   delete mPriv->alphabet;
   delete mPriv->scorer;
 
diff --git a/taskcluster/.build.yml b/taskcluster/.build.yml
index 0cd4385b..bdec4b89 100644
--- a/taskcluster/.build.yml
+++ b/taskcluster/.build.yml
@@ -30,3 +30,4 @@ build:
   args:
     tests_cmdline: ''
   deepspeech_pkg_name: ''
+  convert_graphdef: ''
diff --git a/taskcluster/darwin-amd64-cpu-aot_prod-opt.yml b/taskcluster/darwin-amd64-cpu-aot_prod-opt.yml
index f5210024..cb430482 100644
--- a/taskcluster/darwin-amd64-cpu-aot_prod-opt.yml
+++ b/taskcluster/darwin-amd64-cpu-aot_prod-opt.yml
@@ -4,8 +4,8 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.osx_aot"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.${event.head.sha}.osx_aot"
     - "index.project.deepspeech.deepspeech.native_client.osx_aot.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.osx/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.osx/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.osx/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.osx/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/host-build.sh --aot"
     package: "taskcluster/package.sh"
diff --git a/taskcluster/darwin-amd64-cpu-opt.yml b/taskcluster/darwin-amd64-cpu-opt.yml
index 0c3f3983..50bc7778 100644
--- a/taskcluster/darwin-amd64-cpu-opt.yml
+++ b/taskcluster/darwin-amd64-cpu-opt.yml
@@ -6,8 +6,8 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.osx.${event.head.sha}"
     - "notify.irc-channel.${notifications.irc}.on-exception"
     - "notify.irc-channel.${notifications.irc}.on-failed"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.osx/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.osx/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.osx/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.osx/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/host-build.sh"
     package: "taskcluster/package.sh"
diff --git a/taskcluster/linux-amd64-cpu-aot_prod-opt.yml b/taskcluster/linux-amd64-cpu-aot_prod-opt.yml
index 8e74c4de..a8d490ee 100644
--- a/taskcluster/linux-amd64-cpu-aot_prod-opt.yml
+++ b/taskcluster/linux-amd64-cpu-aot_prod-opt.yml
@@ -4,8 +4,8 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.cpu_aot"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.${event.head.sha}.cpu_aot"
     - "index.project.deepspeech.deepspeech.native_client.cpu_aot.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   system_setup:
     >
       ${nodejs.packages.prep_6} && apt-get -qq update && apt-get -qq -y install nodejs python-yaml &&
diff --git a/taskcluster/linux-amd64-cpu-aot_test-opt.yml b/taskcluster/linux-amd64-cpu-aot_test-opt.yml
index e3abd81f..f68ebbf9 100644
--- a/taskcluster/linux-amd64-cpu-aot_test-opt.yml
+++ b/taskcluster/linux-amd64-cpu-aot_test-opt.yml
@@ -6,8 +6,8 @@ build:
   template_file: linux-opt-base.tyml
   dependencies:
     - "test-training_upstream-linux-amd64-py27-opt"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   system_setup:
     >
       ${nodejs.packages.prep_6} && apt-get -qq update && apt-get -qq -y install nodejs python-yaml &&
diff --git a/taskcluster/linux-amd64-cpu-opt.yml b/taskcluster/linux-amd64-cpu-opt.yml
index bcc25b04..7764dc3b 100644
--- a/taskcluster/linux-amd64-cpu-opt.yml
+++ b/taskcluster/linux-amd64-cpu-opt.yml
@@ -13,8 +13,8 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/host-build.sh"
     package: "taskcluster/package.sh"
diff --git a/taskcluster/linux-amd64-ctc-opt.yml b/taskcluster/linux-amd64-ctc-opt.yml
index 281d06cc..22c8fea5 100644
--- a/taskcluster/linux-amd64-ctc-opt.yml
+++ b/taskcluster/linux-amd64-ctc-opt.yml
@@ -4,8 +4,8 @@ build:
     - "pull_request.synchronize"
     - "pull_request.reopened"
   template_file: linux-opt-base.tyml
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   scripts:
     build: 'taskcluster/decoder-build.sh'
     package: 'taskcluster/decoder-package.sh'
diff --git a/taskcluster/linux-amd64-gpu-opt.yml b/taskcluster/linux-amd64-gpu-opt.yml
index ae4d2446..c42436aa 100644
--- a/taskcluster/linux-amd64-gpu-opt.yml
+++ b/taskcluster/linux-amd64-gpu-opt.yml
@@ -11,8 +11,8 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.gpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.gpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.gpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.gpu/artifacts/public/summarize_graph"
   maxRunTime: 14400
   scripts:
     build: "taskcluster/cuda-build.sh"
diff --git a/taskcluster/linux-rpi3-cpu-aot_prod-opt.yml b/taskcluster/linux-rpi3-cpu-aot_prod-opt.yml
index 424c7fc2..731a0271 100644
--- a/taskcluster/linux-rpi3-cpu-aot_prod-opt.yml
+++ b/taskcluster/linux-rpi3-cpu-aot_prod-opt.yml
@@ -4,8 +4,8 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.arm_aot"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.${event.head.sha}.arm_aot"
     - "index.project.deepspeech.deepspeech.native_client.arm_aot.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.arm/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.arm/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   ## multistrap 2.2.0-ubuntu1 is broken in 14.04: https://bugs.launchpad.net/ubuntu/+source/multistrap/+bug/1313787
   system_setup:
     >
diff --git a/taskcluster/linux-rpi3-cpu-opt.yml b/taskcluster/linux-rpi3-cpu-opt.yml
index 16f9ac2e..255c6bb2 100644
--- a/taskcluster/linux-rpi3-cpu-opt.yml
+++ b/taskcluster/linux-rpi3-cpu-opt.yml
@@ -4,8 +4,8 @@ build:
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.arm"
     - "index.project.deepspeech.deepspeech.native_client.${event.head.branch}.${event.head.sha}.arm"
     - "index.project.deepspeech.deepspeech.native_client.arm.${event.head.sha}"
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.arm/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.arm/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   ## multistrap 2.2.0-ubuntu1 is broken in 14.04: https://bugs.launchpad.net/ubuntu/+source/multistrap/+bug/1313787
   system_setup:
     >
diff --git a/taskcluster/node-package.yml b/taskcluster/node-package.yml
index b1c95949..7da723b9 100644
--- a/taskcluster/node-package.yml
+++ b/taskcluster/node-package.yml
@@ -14,8 +14,8 @@ build:
   system_config:
     >
       ${swig.patch_nodejs.linux}
-  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/home.tar.xz"
-  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.723fb3984980c994579c055736688f693e38ee27.cpu/artifacts/public/summarize_graph"
+  tensorflow: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/home.tar.xz"
+  summarize_graph: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/summarize_graph"
   scripts:
     build: "taskcluster/node-build.sh"
     package: "taskcluster/node-package.sh"
diff --git a/taskcluster/test-darwin-opt-base.tyml b/taskcluster/test-darwin-opt-base.tyml
index bfc51675..005a9bc4 100644
--- a/taskcluster/test-darwin-opt-base.tyml
+++ b/taskcluster/test-darwin-opt-base.tyml
@@ -37,7 +37,8 @@ then:
         DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${darwin_amd64_build}/artifacts/public
         DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package}/artifacts/public
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
-        DEEPSPEECH_PROD_MODEL: https://s3.amazonaws.com/deep-speech/output_graph.pb
+        DEEPSPEECH_PROD_MODEL: https://s3-us-west-2.amazonaws.com/deepspeech/mmap/output_graph.pb
+        DEEPSPEECH_PROD_MODEL_MMAP: https://s3-us-west-2.amazonaws.com/deepspeech/mmap/output_graph.pbmm
 
     command:
       - - "/bin/bash"
diff --git a/taskcluster/test-linux-opt-base.tyml b/taskcluster/test-linux-opt-base.tyml
index 9e527460..84d36921 100644
--- a/taskcluster/test-linux-opt-base.tyml
+++ b/taskcluster/test-linux-opt-base.tyml
@@ -37,13 +37,15 @@ then:
         linux_amd64_ctc: { $eval: as_slugid("linux-amd64-ctc-opt") }
         node_package: { $eval: as_slugid("node-package") }
       in:
+        CONVERT_GRAPHDEF_MEMMAPPED: ${build.convert_graphdef}
         DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_amd64_build}/artifacts/public
         DEEPSPEECH_PYTHON_PACKAGE: https://queue.taskcluster.net/v1/task/${linux_amd64_build}/artifacts/public/${build.deepspeech_pkg_name}
         DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package}/artifacts/public
         DEEPSPEECH_AOT_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_amd64_aot_test}/artifacts/public
         DEEPSPEECH_LIBCTC: https://queue.taskcluster.net/v1/task/${linux_amd64_ctc}/artifacts/public/decoder.tar.xz
         DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
-        DEEPSPEECH_PROD_MODEL: https://s3.amazonaws.com/deep-speech/output_graph.pb
+        DEEPSPEECH_PROD_MODEL: https://s3-us-west-2.amazonaws.com/deepspeech/mmap/output_graph.pb
+        DEEPSPEECH_PROD_MODEL_MMAP: https://s3-us-west-2.amazonaws.com/deepspeech/mmap/output_graph.pbmm
 
     command:
       - "/bin/bash"
diff --git a/taskcluster/test-training_upstream-linux-amd64-py27-opt.yml b/taskcluster/test-training_upstream-linux-amd64-py27-opt.yml
index 288b0e36..fea7202b 100644
--- a/taskcluster/test-training_upstream-linux-amd64-py27-opt.yml
+++ b/taskcluster/test-training_upstream-linux-amd64-py27-opt.yml
@@ -7,6 +7,7 @@ build:
       apt-get -qq -y install ${python.packages.apt}
   args:
     tests_cmdline: "${system.homedir.linux}/DeepSpeech/ds/tc-train-tests.sh 2.7.13 upstream"
+  convert_graphdef: "https://index.taskcluster.net/v1/task/project.deepspeech.tensorflow.pip.r1.5.ad8f785459e80823a2ff4456eeb9d7220c33b9c6.cpu/artifacts/public/convert_graphdef_memmapped_format"
   metadata:
     name: "DeepSpeech Linux AMD64 CPU upstream training Py2.7"
     description: "Training a DeepSpeech LDC93S1 model for Linux/AMD64 using upstream TensorFlow Python 2.7, CPU only, optimized version"
diff --git a/tc-cpp-ds-tests-prod.sh b/tc-cpp-ds-tests-prod.sh
index aa237b94..a7450cdf 100644
--- a/tc-cpp-ds-tests-prod.sh
+++ b/tc-cpp-ds-tests-prod.sh
@@ -7,6 +7,9 @@ source $(dirname "$0")/tc-tests-utils.sh
 model_source=${DEEPSPEECH_PROD_MODEL}
 model_name=$(basename "${model_source}")
 
+model_source_mmap=${DEEPSPEECH_PROD_MODEL_MMAP}
+model_name_mmap=$(basename "${model_source_mmap}")
+
 download_material "${TASKCLUSTER_TMP_DIR}/ds"
 
 export PATH=${TASKCLUSTER_TMP_DIR}/ds/:$PATH
diff --git a/tc-node-tests-prod.sh b/tc-node-tests-prod.sh
index 05c760b4..98f9ba9c 100644
--- a/tc-node-tests-prod.sh
+++ b/tc-node-tests-prod.sh
@@ -14,6 +14,9 @@ fi;
 model_source=${DEEPSPEECH_PROD_MODEL}
 model_name=$(basename "${model_source}")
 
+model_source_mmap=${DEEPSPEECH_PROD_MODEL_MMAP}
+model_name_mmap=$(basename "${model_source_mmap}")
+
 download_data
 
 node --version
diff --git a/tc-python-tests-prod.sh b/tc-python-tests-prod.sh
index e9a48696..74dd3b3c 100644
--- a/tc-python-tests-prod.sh
+++ b/tc-python-tests-prod.sh
@@ -31,6 +31,9 @@ mkdir -p ${PYENV_ROOT} || true
 model_source=${DEEPSPEECH_PROD_MODEL}
 model_name=$(basename "${model_source}")
 
+model_source_mmap=${DEEPSPEECH_PROD_MODEL_MMAP}
+model_name_mmap=$(basename "${model_source_mmap}")
+
 download_data
 
 install_pyenv "${PYENV_ROOT}"
diff --git a/tc-tests-utils.sh b/tc-tests-utils.sh
index d0f49498..34fb1d0f 100755
--- a/tc-tests-utils.sh
+++ b/tc-tests-utils.sh
@@ -37,8 +37,10 @@ export BAZEL_AOT_TARGETS="
 //native_client:libdeepspeech_model.so
 "
 
-model_source=${DEEPSPEECH_TEST_MODEL}
-model_name=$(basename "${model_source}")
+model_source="${DEEPSPEECH_TEST_MODEL}"
+model_name="$(basename "${model_source}")"
+model_name_mmap="$(basename -s ".pb" "${model_source}").pbmm"
+model_source_mmap="$(dirname "${model_source}")/${model_name_mmap}"
 
 SUPPORTED_PYTHON_VERSIONS=${SUPPORTED_PYTHON_VERSIONS:-2.7.13 3.4.6 3.5.3 3.6.2}
 SUPPORTED_NODEJS_VERSIONS=${SUPPORTED_NODEJS_VERSIONS:-4.8.6 5.12.0 6.12.0 7.10.1 8.9.1 9.2.0}
@@ -147,14 +149,19 @@ assert_correct_multi_ldc93s1()
   # assert_shows_something "$1" "/LDC93S1_pcms16le_1_8000.wav%she hayorasryrtl lyreasy asr watal w water all year%"
 }
 
-assert_correct_ldc93s1_prodmodel()
+assert_correct_ldc93s1_prodmodel_v1()
 {
   assert_correct_inference "$1" "she had the duck so ingrecywachworallyear"
 }
 
+assert_correct_ldc93s1_prodmodel_v2()
+{
+  assert_correct_inference "$1" "she had a ducsuotangresywathorerall year"
+}
+
 assert_working_ldc93s1_prodmodel()
 {
-  assert_working_inference "$1" "she had the duck so"
+  assert_working_inference "$1" "she had"
 }
 
 assert_correct_ldc93s1_somodel()
@@ -198,19 +205,22 @@ run_all_inference_tests()
   phrase_pbmodel_nolm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
   assert_correct_ldc93s1 "${phrase_pbmodel_nolm}"
 
-  phrase_pbmodel_withlm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
+  phrase_pbmodel_nolm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
+  assert_correct_ldc93s1 "${phrase_pbmodel_nolm}"
+
+  phrase_pbmodel_withlm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
   assert_correct_ldc93s1 "${phrase_pbmodel_withlm}"
 
-  phrase_pbmodel_nolm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
+  phrase_pbmodel_nolm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
   assert_correct_ldc93s1 "${phrase_pbmodel_nolm_stereo_44k}"
 
-  phrase_pbmodel_withlm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
+  phrase_pbmodel_withlm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
   assert_correct_ldc93s1 "${phrase_pbmodel_withlm_stereo_44k}"
 
-  phrase_pbmodel_nolm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
+  phrase_pbmodel_nolm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
   assert_correct_warning_upsampling "${phrase_pbmodel_nolm_mono_8k}"
 
-  phrase_pbmodel_withlm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
+  phrase_pbmodel_withlm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
   assert_correct_warning_upsampling "${phrase_pbmodel_withlm_mono_8k}"
 
   if [ "${aot_model}" = "--aot" ]; then
@@ -234,12 +244,15 @@ run_all_inference_tests()
 run_prod_inference_tests()
 {
   phrase_pbmodel_withlm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
-  assert_correct_ldc93s1_prodmodel "${phrase_pbmodel_withlm}"
+  assert_correct_ldc93s1_prodmodel_v1 "${phrase_pbmodel_withlm}"
 
-  phrase_pbmodel_withlm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
+  phrase_pbmodel_withlm=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1.wav)
+  assert_correct_ldc93s1_prodmodel_v2 "${phrase_pbmodel_withlm}"
+
+  phrase_pbmodel_withlm_stereo_44k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_2_44100.wav)
   assert_working_ldc93s1_prodmodel "${phrase_pbmodel_withlm_stereo_44k}"
 
-  phrase_pbmodel_withlm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
+  phrase_pbmodel_withlm_mono_8k=$(deepspeech ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/alphabet.txt ${TASKCLUSTER_TMP_DIR}/lm.binary ${TASKCLUSTER_TMP_DIR}/trie ${TASKCLUSTER_TMP_DIR}/LDC93S1_pcms16le_1_8000.wav 2>&1 1>/dev/null)
   assert_correct_warning_upsampling "${phrase_pbmodel_withlm_mono_8k}"
 }
 
@@ -286,7 +299,8 @@ download_ctc_kenlm()
 
 download_data()
 {
-  wget ${model_source} -O ${TASKCLUSTER_TMP_DIR}/${model_name}
+  wget -P "${TASKCLUSTER_TMP_DIR}" "${model_source}"
+  wget -P "${TASKCLUSTER_TMP_DIR}" "${model_source_mmap}"
   cp ${DS_ROOT_TASK}/DeepSpeech/ds/data/smoke_test/*.wav ${TASKCLUSTER_TMP_DIR}/
   cp ${DS_ROOT_TASK}/DeepSpeech/ds/data/alphabet.txt ${TASKCLUSTER_TMP_DIR}/alphabet.txt
   cp ${DS_ROOT_TASK}/DeepSpeech/ds/data/lm/lm.binary ${TASKCLUSTER_TMP_DIR}/lm.binary
@@ -306,7 +320,7 @@ download_material()
 
   download_data
 
-  ls -hal ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/LDC93S1*.wav ${TASKCLUSTER_TMP_DIR}/alphabet.txt
+  ls -hal ${TASKCLUSTER_TMP_DIR}/${model_name} ${TASKCLUSTER_TMP_DIR}/${model_name_mmap} ${TASKCLUSTER_TMP_DIR}/LDC93S1*.wav ${TASKCLUSTER_TMP_DIR}/alphabet.txt
 }
 
 install_pyenv()
@@ -351,9 +365,7 @@ do_get_model_parameters()
   fi;
 
   wget "${model_url}" -O "${model_file}"
-  wget "${SUMMARIZE_GRAPH_BINARY}" -O "/tmp/summarize_graph"
-
-  chmod +x /tmp/summarize_graph
+  wget -P "/tmp/" "${SUMMARIZE_GRAPH_BINARY}" && chmod +x /tmp/summarize_graph
 
   if [ ! -f "${model_file}" ]; then
     echo "No such model: ${model_file}"
diff --git a/tc-train-tests.sh b/tc-train-tests.sh
index e5b89690..0a208446 100644
--- a/tc-train-tests.sh
+++ b/tc-train-tests.sh
@@ -60,3 +60,11 @@ deactivate
 pyenv uninstall --force ${PYENV_NAME}
 
 cp /tmp/train/output_graph.pb ${TASKCLUSTER_ARTIFACTS}
+
+if [ ! -z "${CONVERT_GRAPHDEF_MEMMAPPED}" ]; then
+  convert_graphdef=$(basename "${CONVERT_GRAPHDEF_MEMMAPPED}")
+  wget -P "/tmp/" "${CONVERT_GRAPHDEF_MEMMAPPED}" && chmod +x "/tmp/${convert_graphdef}"
+
+  /tmp/${convert_graphdef} --in_graph=/tmp/train/output_graph.pb --out_graph=/tmp/train/output_graph.pbmm
+  cp /tmp/train/output_graph.pbmm ${TASKCLUSTER_ARTIFACTS}
+fi;