diff --git a/examples/models/llava/export_llava.py b/examples/models/llava/export_llava.py
index d95bd7fb054..9ba6a510736 100644
--- a/examples/models/llava/export_llava.py
+++ b/examples/models/llava/export_llava.py
@@ -226,11 +226,11 @@ def export_all(llava_model: LlavaModel):
         {
             "image_encoder": image_encoder_ep,
             "token_embedding": token_embedding_ep,
-            "text_model": text_model_ep,
+            "text_decoder": text_model_ep,
         },
         partitioner={
             "image_encoder": [XnnpackPartitioner()],
-            "text_model": [
+            "text_decoder": [
                 # First partition the DQLinear nodes, then partition the rest of the nodes,
                 # to avoid multiple DQLinear nodes in the same partition,
                 # to avoid holding multiple unpacked and packed weight buffers in memory,
@@ -254,7 +254,7 @@ def export_all(llava_model: LlavaModel):
             memory_planning_pass=MemoryPlanningPass(alloc_graph_input=False),
             sym_shape_eval_pass={
                 "image_encoder": ConstraintBasedSymShapeEvalPass(),
-                "text_model": ConstraintBasedSymShapeEvalPass(),
+                "text_decoder": ConstraintBasedSymShapeEvalPass(),
                 "token_embedding": HintBasedSymShapeEvalPass(),
             },
         )
diff --git a/examples/models/llava/runner/llava_text_decoder_runner.h b/examples/models/llava/runner/llava_text_decoder_runner.h
index 09b8e82d49d..cfa92e0c253 100644
--- a/examples/models/llava/runner/llava_text_decoder_runner.h
+++ b/examples/models/llava/runner/llava_text_decoder_runner.h
@@ -89,7 +89,7 @@ class ET_EXPERIMENTAL LlavaTextDecoderRunner
   }
 
   inline static const std::string kTokenEmbeddingMethod = "token_embedding";
-  inline static const std::string kTextModelMethod = "text_model";
+  inline static const std::string kTextModelMethod = "text_decoder";
 };
 
 } // namespace example
diff --git a/examples/models/llava/test/test_llava.py b/examples/models/llava/test/test_llava.py
index 05cfd5b1497..def9eaa02bd 100644
--- a/examples/models/llava/test/test_llava.py
+++ b/examples/models/llava/test/test_llava.py
@@ -96,7 +96,7 @@ def test_llava_export(self):
             "token_embedding", (prompt_before_image,)
         )[0]
         llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_before_img),
         )
 
@@ -107,7 +107,7 @@ def test_llava_export(self):
         # pte prefill image
         pte_embeds_img = llava_module.run_method("image_encoder", (resized,))[0]
         llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (
                 torch.tensor([start_pos], dtype=torch.int64),
                 pte_embeds_img,
@@ -122,7 +122,7 @@ def test_llava_export(self):
             "token_embedding", (prompt_after_image,)
         )[0]
         pte_prefill_after_img = llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_after_img),
         )[0]
 
@@ -139,7 +139,7 @@ def test_llava_export(self):
                 "token_embedding", (torch.tensor([[new_tokens[i]]], dtype=torch.int64),)
             )[0]
             logits = llava_module.run_method(
-                "text_model",
+                "text_decoder",
                 (torch.tensor([start_pos + i], dtype=torch.int64), token_embeds),
             )[0]
             new_tokens.append(torch.argmax(logits).item())
diff --git a/examples/models/llava/test/test_pte.py b/examples/models/llava/test/test_pte.py
index f12d72f854b..1f4aaa9938c 100644
--- a/examples/models/llava/test/test_pte.py
+++ b/examples/models/llava/test/test_pte.py
@@ -47,7 +47,7 @@ def main():
         "token_embedding", (prompt_before_image,)
     )[0]
     pte_prefill_before_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_before_img),
     )[0]
     print(pte_prefill_before_img)
@@ -60,7 +60,7 @@ def main():
     logging.warning("Image encoder finished")
     logging.warning("Image token prefill started")
     pte_prefill_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (
             torch.tensor([start_pos], dtype=torch.int64),
             pte_embeds_img,
@@ -77,7 +77,7 @@ def main():
         "token_embedding", (prompt_after_image,)
     )[0]
     pte_prefill_after_img = llava_module.run_method(
-        "text_model",
+        "text_decoder",
         (torch.tensor([start_pos], dtype=torch.int64), pte_embeds_after_img),
     )[0]
     logging.warning("Text token prefill finished")
@@ -91,7 +91,7 @@ def main():
             "token_embedding", (torch.tensor([[new_tokens[i]]], dtype=torch.int64),)
         )[0]
         logits = llava_module.run_method(
-            "text_model",
+            "text_decoder",
             (torch.tensor([start_pos + i], dtype=torch.int64), token_embeds),
         )[0]
         new_tokens.append(torch.argmax(logits[..., -1, :]).item())
diff --git a/extension/llm/runner/audio.h b/extension/llm/runner/audio.h
new file mode 100644
index 00000000000..868765950af
--- /dev/null
+++ b/extension/llm/runner/audio.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) Meta Platforms, Inc. and affiliates.
+ * All rights reserved.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+// A simple audio struct.
+
+#pragma once
+#include <executorch/runtime/platform/compiler.h>
+#include <cstdint>
+#include <vector>
+
+namespace executorch {
+namespace extension {
+namespace llm {
+
+/**
+ * Audio inputs as a raw audio tensor, for use when the audio processing
+ * into a mel spectrogram is baked into the audio encoder with torch.export.
+ */
+struct ET_EXPERIMENTAL RawAudio {
+  std::vector<uint8_t> data;
+  int32_t batch_size;
+  int32_t n_channels; // For mono, use n_channels = 1.
+  int32_t n_samples;
+};
+
+/**
+ * Pre-processed audio inputs, ready to feed directly into an audio
+ * encoder.
+ */
+struct ET_EXPERIMENTAL Audio {
+  std::vector<uint8_t> data;
+  int32_t batch_size;
+  int32_t n_bins;
+  int32_t n_frames;
+};
+
+} // namespace llm
+} // namespace extension
+} // namespace executorch
+
+namespace torch {
+namespace executor {
+// TODO(T197294990): Remove these deprecated aliases once all users have moved
+// to the new `::executorch` namespaces.
+using ::executorch::extension::llm::Audio;
+} // namespace executor
+} // namespace torch
diff --git a/extension/llm/runner/constants.h b/extension/llm/runner/constants.h
index fc6ddcb451c..4ba88203c50 100644
--- a/extension/llm/runner/constants.h
+++ b/extension/llm/runner/constants.h
@@ -21,7 +21,8 @@ inline constexpr auto kUseSDPAWithKVCache = "use_sdpa_with_kv_cache";
 
 // Multimodal method name conventions
 inline constexpr auto kImageEncoderMethod = "image_encoder";
+inline constexpr auto kAudioEncoderMethod = "audio_encoder";
 inline constexpr auto kTokenEmbeddingMethod = "token_embedding";
-inline constexpr auto kTextModelMethod = "text_model";
+inline constexpr auto kTextModelMethod = "text_decoder";
 
 } // namespace executorch::extension::llm
diff --git a/extension/llm/runner/multimodal_input.h b/extension/llm/runner/multimodal_input.h
index ae243992fec..728d8aef08f 100644
--- a/extension/llm/runner/multimodal_input.h
+++ b/extension/llm/runner/multimodal_input.h
@@ -11,6 +11,7 @@
 
 #pragma once
 
+#include <executorch/extension/llm/runner/audio.h>
 #include <executorch/extension/llm/runner/image.h>
 #include <executorch/runtime/platform/compiler.h>
 #include <string>
@@ -19,19 +20,31 @@
 namespace executorch::extension::llm {
 
 /**
- * A generic class to hold either image or text data for multimodal inputs.
- * This allows the generate() API to take a std::vector of these objects
- * instead of separate image and text parameters.
+ * A generic class to hold either image, text, or audio data for multimodal
+ * inputs. This allows the generate() API to take a std::vector of these objects
+ * instead of separate image, text, and audio parameters.
  */
 class ET_EXPERIMENTAL MultimodalInput {
  public:
-  enum class Type { TEXT, IMAGE };
+  /// Type of multimodal input data
+  enum class Type {
+    TEXT, ///< Text string input
+    IMAGE, ///< Processed image input
+    AUDIO, ///< Processed audio input
+    RAW_AUDIO, ///< Raw unprocessed audio input (straight from audio file)
+    UNSUPPORTED ///< Unsupported input type
+  };
 
   // Constructors
   explicit MultimodalInput(const std::string& text) : data_(text) {}
   explicit MultimodalInput(std::string&& text) : data_(std::move(text)) {}
   explicit MultimodalInput(const Image& image) : data_(image) {}
   explicit MultimodalInput(Image&& image) : data_(std::move(image)) {}
+  explicit MultimodalInput(const Audio& audio) : data_(audio) {}
+  explicit MultimodalInput(Audio&& audio) : data_(std::move(audio)) {}
+  explicit MultimodalInput(const RawAudio& raw_audio) : data_(raw_audio) {}
+  explicit MultimodalInput(RawAudio&& raw_audio)
+      : data_(std::move(raw_audio)) {}
 
   // Copy constructor and assignment
   MultimodalInput(const MultimodalInput& other) = default;
@@ -60,12 +73,37 @@ class ET_EXPERIMENTAL MultimodalInput {
     return std::holds_alternative<Image>(data_);
   }
 
+  /**
+   * Check if this input contains audio data.
+   * @return true if this input contains audio, false otherwise.
+   */
+  bool is_audio() const noexcept {
+    return std::holds_alternative<Audio>(data_);
+  }
+
+  /**
+   * Check if this input contains raw audio data.
+   * @return true if this input contains raw audio, false otherwise.
+   */
+  bool is_raw_audio() const noexcept {
+    return std::holds_alternative<RawAudio>(data_);
+  }
+
   /**
    * Get the type of data stored in this input.
-   * @return Type::TEXT if text data, Type::IMAGE if image data.
+   * @return Type::TEXT if text data, Type::IMAGE if image data, Type::AUDIO if
+   * audio data, Type::RAW_AUDIO if raw audio data.
    */
   Type get_type() const noexcept {
-    return is_text() ? Type::TEXT : Type::IMAGE;
+    if (is_text())
+      return Type::TEXT;
+    if (is_image())
+      return Type::IMAGE;
+    if (is_audio())
+      return Type::AUDIO;
+    if (is_raw_audio())
+      return Type::RAW_AUDIO;
+    return Type::UNSUPPORTED;
   }
 
   /**
@@ -122,6 +160,60 @@ class ET_EXPERIMENTAL MultimodalInput {
     return std::get<Image>(std::move(data_));
   }
 
+  /**
+   * Get the audio data from this input.
+   * @return Reference to the stored Audio object.
+   * @throws std::bad_variant_access if this input doesn't contain audio.
+   */
+  const Audio& get_audio() const& {
+    return std::get<Audio>(data_);
+  }
+
+  /**
+   * Get the audio data from this input (mutable version).
+   * @return Mutable reference to the stored Audio object.
+   * @throws std::bad_variant_access if this input doesn't contain audio.
+   */
+  Audio& get_audio() & {
+    return std::get<Audio>(data_);
+  }
+
+  /**
+   * Get the audio data from this input (rvalue version).
+   * @return Rvalue reference to the stored Audio object for efficient moves.
+   * @throws std::bad_variant_access if this input doesn't contain audio.
+   */
+  Audio&& get_audio() && {
+    return std::get<Audio>(std::move(data_));
+  }
+
+  /**
+   * Get the raw audio data from this input.
+   * @return Reference to the stored RawAudio object.
+   * @throws std::bad_variant_access if this input doesn't contain raw audio.
+   */
+  const RawAudio& get_raw_audio() const& {
+    return std::get<RawAudio>(data_);
+  }
+
+  /**
+   * Get the raw audio data from this input (mutable version).
+   * @return Mutable reference to the stored RawAudio object.
+   * @throws std::bad_variant_access if this input doesn't contain raw audio.
+   */
+  RawAudio& get_raw_audio() & {
+    return std::get<RawAudio>(data_);
+  }
+
+  /**
+   * Get the raw audio data from this input (rvalue version).
+   * @return Rvalue reference to the stored RawAudio object for efficient moves.
+   * @throws std::bad_variant_access if this input doesn't contain raw audio.
+   */
+  RawAudio&& get_raw_audio() && {
+    return std::get<RawAudio>(std::move(data_));
+  }
+
   /**
    * Try to get the text data from this input safely.
    * @return Pointer to the text string if this input contains text, nullptr
@@ -158,8 +250,44 @@ class ET_EXPERIMENTAL MultimodalInput {
     return std::get_if<Image>(&data_);
   }
 
+  /**
+   * Try to get the audio data from this input safely.
+   * @return Pointer to the Audio object if this input contains audio,
+   * nullptr otherwise.
+   */
+  const Audio* try_get_audio() const noexcept {
+    return std::get_if<Audio>(&data_);
+  }
+
+  /**
+   * Try to get the audio data from this input safely (mutable version).
+   * @return Pointer to the Audio object if this input contains audio,
+   * nullptr otherwise.
+   */
+  Audio* try_get_audio() noexcept {
+    return std::get_if<Audio>(&data_);
+  }
+
+  /**
+   * Try to get the raw audio data from this input safely.
+   * @return Pointer to the RawAudio object if this input contains raw audio,
+   * nullptr otherwise.
+   */
+  const RawAudio* try_get_raw_audio() const noexcept {
+    return std::get_if<RawAudio>(&data_);
+  }
+
+  /**
+   * Try to get the raw audio data from this input safely (mutable version).
+   * @return Pointer to the RawAudio object if this input contains raw audio,
+   * nullptr otherwise.
+   */
+  RawAudio* try_get_raw_audio() noexcept {
+    return std::get_if<RawAudio>(&data_);
+  }
+
  private:
-  std::variant<std::string, Image> data_;
+  std::variant<std::string, Image, Audio, RawAudio> data_;
 };
 
 // Convenience factory functions
@@ -179,4 +307,21 @@ inline MultimodalInput make_image_input(Image&& image) noexcept {
   return MultimodalInput(std::move(image));
 }
 
-} // namespace executorch::extension::llm
\ No newline at end of file
+inline MultimodalInput make_audio_input(const Audio& audio) noexcept {
+  return MultimodalInput(audio);
+}
+
+inline MultimodalInput make_audio_input(Audio&& audio) noexcept {
+  return MultimodalInput(std::move(audio));
+}
+
+inline MultimodalInput make_raw_audio_input(
+    const RawAudio& raw_audio) noexcept {
+  return MultimodalInput(raw_audio);
+}
+
+inline MultimodalInput make_raw_audio_input(RawAudio&& raw_audio) noexcept {
+  return MultimodalInput(std::move(raw_audio));
+}
+
+} // namespace executorch::extension::llm
diff --git a/extension/llm/runner/multimodal_prefiller.cpp b/extension/llm/runner/multimodal_prefiller.cpp
index 7f69041551f..1d9a0c8fdfc 100644
--- a/extension/llm/runner/multimodal_prefiller.cpp
+++ b/extension/llm/runner/multimodal_prefiller.cpp
@@ -37,7 +37,7 @@ MultimodalPrefiller::MultimodalPrefiller(
 Result<uint64_t> MultimodalPrefiller::prefill(
     const MultimodalInput& input,
     int64_t& start_pos) {
-  // Check if input is image
+  // 1. Run encoder model.
   ::executorch::runtime::EValue encoder_output;
   if (input.is_image()) {
     Image image = input.get_image();
@@ -51,34 +51,79 @@ Result<uint64_t> MultimodalPrefiller::prefill(
         ET_UNWRAP(module_->execute(kImageEncoderMethod, image_tensor));
 
     encoder_output = image_encoder_outputs[0];
+  } else if (input.is_audio()) {
+    Audio audio = input.get_audio();
+
+    // Use the original tensor shape as intended
+    auto audio_tensor = executorch::extension::from_blob(
+        audio.data.data(),
+        {audio.batch_size, audio.n_bins, audio.n_frames},
+        ::executorch::aten::ScalarType::Float);
+
+    // Run audio encoder
+    auto audio_encoder_result =
+        module_->execute(kAudioEncoderMethod, audio_tensor);
+    if (audio_encoder_result.error() != ::executorch::runtime::Error::Ok) {
+      return ::executorch::runtime::Error::Internal;
+    }
+    auto audio_encoder_outputs = audio_encoder_result.get();
+
+    encoder_output = audio_encoder_outputs[0];
   } else if (input.is_text()) {
-    // For text input, we don't need to run the image encoder.
-    // Instead, we run the text encoder to get the encoder output.
     auto& text = input.get_text();
     std::vector<uint64_t> tokens =
         ET_UNWRAP_TOKENIZER(tokenizer_->encode(text));
+
     auto text_tensor = executorch::extension::from_blob(
         tokens.data(),
         {1, static_cast<aten::SizesType>(tokens.size())},
         ::executorch::aten::ScalarType::Long);
 
-    // Run token embedding
+    // Run text encoder (token embeddings)
     auto token_embedding_outputs =
         ET_UNWRAP(module_->execute(kTokenEmbeddingMethod, text_tensor));
 
     encoder_output = token_embedding_outputs[0];
   } else {
     ET_LOG(Error, "Unsupported input type");
-    // For all other input types (e.g., audio), return error
+    // For any other input types, return error
     return ::executorch::runtime::Error::NotSupported;
   }
 
-  auto outputs_res =
-      ET_UNWRAP(text_decoder_runner_->decode(encoder_output, start_pos));
+  // 2. Run decoder model for prefill.
+  // `cache_position` goes from start_pos to start_pos + encoder_output.size(1).
+  // e.g. if start_pos = 2 and encoder_output.size(1) = 5,
+  // cache_position_tensor should be [2, 3, 4, 5, 6].
+  int64_t seq_len = encoder_output.toTensor().size(1);
+  if (seq_len == 0) {
+    ET_LOG(Error, "The encoder returned an empty output.");
+    return ::executorch::runtime::Error::InvalidState;
+  }
+  std::vector<int64_t> cache_positions(seq_len);
+  for (int64_t i = 0; i < seq_len; ++i) {
+    cache_positions[i] = start_pos + i;
+  }
+  auto cache_position_tensor = ::executorch::extension::from_blob(
+      cache_positions.data(),
+      {static_cast<int>(seq_len)},
+      executorch::aten::ScalarType::Long);
+  auto prefill_result = module_->execute(
+      kTextModelMethod, {cache_position_tensor, encoder_output});
+  if (prefill_result.error() != ::executorch::runtime::Error::Ok) {
+    return prefill_result.error();
+  }
+  // Check if prefill_outputs is empty, if it is return error and log that the
+  // specified encoder returned empty results when used to prefill decoder.
+  auto prefill_outputs = prefill_result.get();
+  if (prefill_outputs.empty()) {
+    ET_LOG(
+        Error, "Encoder returned empty results when used to prefill decoder");
+    return ::executorch::runtime::Error::InvalidState;
+  }
+  auto outputs_res = prefill_outputs[0].toTensor();
 
-  // Update the start_pos, which is only available inside this function.
-  // outputs_res can have only one logits.
-  start_pos += encoder_output.toTensor().size(1);
+  // Update start_pos, tracking the current cache position.
+  start_pos += seq_len;
 
   return static_cast<uint64_t>(
       text_decoder_runner_->logits_to_token(outputs_res));
@@ -103,6 +148,11 @@ ::executorch::runtime::Error MultimodalPrefiller::load() {
   if (methods.find(kImageEncoderMethod) != methods.end()) {
     ET_CHECK_OK_OR_RETURN_ERROR(module_->load_method(kImageEncoderMethod));
   }
+
+  if (methods.find(kAudioEncoderMethod) != methods.end()) {
+    ET_CHECK_OK_OR_RETURN_ERROR(module_->load_method(kAudioEncoderMethod));
+  }
+
   return ::executorch::runtime::Error::Ok;
 }
 
diff --git a/extension/llm/runner/targets.bzl b/extension/llm/runner/targets.bzl
index 05f05ac6fad..8ae0c1ba3be 100644
--- a/extension/llm/runner/targets.bzl
+++ b/extension/llm/runner/targets.bzl
@@ -98,6 +98,8 @@ def define_common_targets():
         runtime.cxx_library(
             name = "multimodal_runner_lib" + aten_suffix,
             exported_headers = [
+                "audio.h",
+                "image.h",
                 "multimodal_input.h",
                 "multimodal_runner.h",
                 "multimodal_prefiller.h",
@@ -143,6 +145,7 @@ def define_common_targets():
                 "//pytorch/tokenizers:hf_tokenizer",
                 "//pytorch/tokenizers:llama2c_tokenizer",
                 "//pytorch/tokenizers:sentencepiece",
+                "//pytorch/tokenizers:tekken",
                 "//pytorch/tokenizers:tiktoken",
             ],
         )