diff --git a/ai-logic/firebase-ai-ondevice-interop/CHANGELOG.md b/ai-logic/firebase-ai-ondevice-interop/CHANGELOG.md
index 18d392d7039..c618d823a53 100644
--- a/ai-logic/firebase-ai-ondevice-interop/CHANGELOG.md
+++ b/ai-logic/firebase-ai-ondevice-interop/CHANGELOG.md
@@ -1,5 +1,7 @@
 # Unreleased
 
+- [feature] Added support for on-device system instructions.
+
 # 16.0.0-beta03
 
 - [feature] Added `modelVersion` property to `GenerateContentResponse` (#8227)
diff --git a/ai-logic/firebase-ai-ondevice-interop/api.txt b/ai-logic/firebase-ai-ondevice-interop/api.txt
index 42f19306399..96efc52b99d 100644
--- a/ai-logic/firebase-ai-ondevice-interop/api.txt
+++ b/ai-logic/firebase-ai-ondevice-interop/api.txt
@@ -79,11 +79,12 @@ package com.google.firebase.ai.ondevice.interop {
   }
 
   public final class GenerateContentRequest {
-    ctor public GenerateContentRequest(com.google.firebase.ai.ondevice.interop.TextPart text, com.google.firebase.ai.ondevice.interop.ImagePart? image = null, Float? temperature = null, Integer? topK = null, Integer? seed = null, Integer? candidateCount = null, Integer? maxOutputTokens = null);
+    ctor public GenerateContentRequest(com.google.firebase.ai.ondevice.interop.TextPart text, com.google.firebase.ai.ondevice.interop.ImagePart? image = null, Float? temperature = null, Integer? topK = null, Integer? seed = null, Integer? candidateCount = null, Integer? maxOutputTokens = null, com.google.firebase.ai.ondevice.interop.TextPart? systemInstruction = null);
     method public Integer? getCandidateCount();
     method public com.google.firebase.ai.ondevice.interop.ImagePart? getImage();
     method public Integer? getMaxOutputTokens();
     method public Integer? getSeed();
+    method public com.google.firebase.ai.ondevice.interop.TextPart? getSystemInstruction();
     method public Float? getTemperature();
     method public com.google.firebase.ai.ondevice.interop.TextPart getText();
     method public Integer? getTopK();
@@ -91,6 +92,7 @@ package com.google.firebase.ai.ondevice.interop {
     property public final com.google.firebase.ai.ondevice.interop.ImagePart? image;
     property public final Integer? maxOutputTokens;
     property public final Integer? seed;
+    property public final com.google.firebase.ai.ondevice.interop.TextPart? systemInstruction;
     property public final Float? temperature;
     property public final com.google.firebase.ai.ondevice.interop.TextPart text;
     property public final Integer? topK;
diff --git a/ai-logic/firebase-ai-ondevice-interop/src/main/kotlin/com/google/firebase/ai/ondevice/interop/GenerateContentRequest.kt b/ai-logic/firebase-ai-ondevice-interop/src/main/kotlin/com/google/firebase/ai/ondevice/interop/GenerateContentRequest.kt
index bfe2877887b..65ae961552d 100644
--- a/ai-logic/firebase-ai-ondevice-interop/src/main/kotlin/com/google/firebase/ai/ondevice/interop/GenerateContentRequest.kt
+++ b/ai-logic/firebase-ai-ondevice-interop/src/main/kotlin/com/google/firebase/ai/ondevice/interop/GenerateContentRequest.kt
@@ -30,6 +30,7 @@ package com.google.firebase.ai.ondevice.interop
  * @property candidateCount The number of candidates to generate.
  * @property maxOutputTokens Specifies the maximum number of tokens that can be generated in the
  * response.
+ * @property systemInstruction The system instructions for directing model behavior.
  */
 public class GenerateContentRequest(
   public val text: TextPart,
@@ -39,4 +40,5 @@ public class GenerateContentRequest(
   public val seed: Int? = null,
   public val candidateCount: Int? = null,
   public val maxOutputTokens: Int? = null,
+  public val systemInstruction: TextPart? = null,
 ) {}
diff --git a/ai-logic/firebase-ai-ondevice/CHANGELOG.md b/ai-logic/firebase-ai-ondevice/CHANGELOG.md
index 04dab651cfd..823a866de61 100644
--- a/ai-logic/firebase-ai-ondevice/CHANGELOG.md
+++ b/ai-logic/firebase-ai-ondevice/CHANGELOG.md
@@ -1,6 +1,7 @@
 # Unreleased
 
 - [feature] Added `getOnDeviceModelName` to `GenerativeModel` (#8247)
+- [feature] Added support for on-device system instructions.
 
 # 16.0.0-beta03
 
diff --git a/ai-logic/firebase-ai-ondevice/src/main/kotlin/com/google/firebase/ai/ondevice/Converters.kt b/ai-logic/firebase-ai-ondevice/src/main/kotlin/com/google/firebase/ai/ondevice/Converters.kt
index 3b3e2454f9a..78fd67d2c4c 100644
--- a/ai-logic/firebase-ai-ondevice/src/main/kotlin/com/google/firebase/ai/ondevice/Converters.kt
+++ b/ai-logic/firebase-ai-ondevice/src/main/kotlin/com/google/firebase/ai/ondevice/Converters.kt
@@ -31,6 +31,7 @@ import com.google.mlkit.genai.prompt.GenerateContentRequest
 import com.google.mlkit.genai.prompt.ImagePart
 import com.google.mlkit.genai.prompt.ModelPreference
 import com.google.mlkit.genai.prompt.ModelReleaseStage
+import com.google.mlkit.genai.prompt.SystemInstruction
 import com.google.mlkit.genai.prompt.TextPart
 import com.google.mlkit.genai.prompt.generationConfig
 import com.google.mlkit.genai.prompt.modelConfig
@@ -74,6 +75,7 @@ internal fun com.google.firebase.ai.ondevice.interop.GenerateContentRequest.toMl
       candidateCount = this@toMlKit.candidateCount
       topK = this@toMlKit.topK
       seed = this@toMlKit.seed
+      this@toMlKit.systemInstruction?.let { systemInstruction = SystemInstruction(it.text) }
     }
   } catch (e: IllegalArgumentException) {
     throw FirebaseAIOnDeviceInvalidRequestException(e)
diff --git a/ai-logic/firebase-ai-ondevice/src/test/kotlin/com/google/firebase/ai/ondevice/ConvertersTest.kt b/ai-logic/firebase-ai-ondevice/src/test/kotlin/com/google/firebase/ai/ondevice/ConvertersTest.kt
index 92c51816ad7..0b4821b4428 100644
--- a/ai-logic/firebase-ai-ondevice/src/test/kotlin/com/google/firebase/ai/ondevice/ConvertersTest.kt
+++ b/ai-logic/firebase-ai-ondevice/src/test/kotlin/com/google/firebase/ai/ondevice/ConvertersTest.kt
@@ -101,7 +101,19 @@ internal class ConvertersTest {
     assertThat(mlKitRequest.topK).isEqualTo(3)
     assertThat(mlKitRequest.seed).isEqualTo(0)
     assertThat(mlKitRequest.candidateCount).isEqualTo(1)
-    assertThat(mlKitRequest.maxOutputTokens).isEqualTo(256)
+    assertThat(mlKitRequest.maxOutputTokens).isEqualTo(4096)
+  }
+
+  @Test
+  fun `GenerateContentRequest toMlKit should convert systemInstruction correctly`() {
+    val interopRequest =
+      InteropGenerateContentRequest(
+        text = InteropTextPart("prompt"),
+        systemInstruction = InteropTextPart("system instruction")
+      )
+    val mlKitRequest = interopRequest.toMlKit()
+
+    assertThat(mlKitRequest.systemInstruction?.textString).isEqualTo("system instruction")
   }
 
   @Test
diff --git a/ai-logic/firebase-ai/CHANGELOG.md b/ai-logic/firebase-ai/CHANGELOG.md
index 3c8437ead69..e2d86f5af52 100644
--- a/ai-logic/firebase-ai/CHANGELOG.md
+++ b/ai-logic/firebase-ai/CHANGELOG.md
@@ -9,6 +9,7 @@
   Note: The default location is now `global` instead of `us-central1` (no other
   functionality has changed). To continue using `us-central1`, specify
   `GenerativeBackend.agentPlatform(location = "us-central1"))`.
+- [feature] Added support for on-device system instructions.
 
 # 17.14.0
 
diff --git a/ai-logic/firebase-ai/api.txt b/ai-logic/firebase-ai/api.txt
index 9b02089f183..995f6724afe 100644
--- a/ai-logic/firebase-ai/api.txt
+++ b/ai-logic/firebase-ai/api.txt
@@ -150,11 +150,13 @@ package com.google.firebase.ai {
     ctor public OnDeviceConfig(com.google.firebase.ai.InferenceMode mode, Integer? maxOutputTokens = null, Float? temperature = null, Integer? topK = null, Integer? seed = null);
     ctor public OnDeviceConfig(com.google.firebase.ai.InferenceMode mode, Integer? maxOutputTokens = null, Float? temperature = null, Integer? topK = null, Integer? seed = null, int candidateCount = 1);
     ctor public OnDeviceConfig(com.google.firebase.ai.InferenceMode mode, Integer? maxOutputTokens = null, Float? temperature = null, Integer? topK = null, Integer? seed = null, int candidateCount = 1, com.google.firebase.ai.OnDeviceModelOption? modelOption = null);
+    ctor public OnDeviceConfig(com.google.firebase.ai.InferenceMode mode, Integer? maxOutputTokens = null, Float? temperature = null, Integer? topK = null, Integer? seed = null, int candidateCount = 1, com.google.firebase.ai.OnDeviceModelOption? modelOption = null, com.google.firebase.ai.type.Content? systemInstruction = null);
     method public int getCandidateCount();
     method public Integer? getMaxOutputTokens();
     method public com.google.firebase.ai.InferenceMode getMode();
     method public com.google.firebase.ai.OnDeviceModelOption? getModelOption();
     method public Integer? getSeed();
+    method public com.google.firebase.ai.type.Content? getSystemInstruction();
     method public Float? getTemperature();
     method public Integer? getTopK();
     property public final int candidateCount;
@@ -162,6 +164,7 @@ package com.google.firebase.ai {
     property public final com.google.firebase.ai.InferenceMode mode;
     property public final com.google.firebase.ai.OnDeviceModelOption? modelOption;
     property public final Integer? seed;
+    property public final com.google.firebase.ai.type.Content? systemInstruction;
     property public final Float? temperature;
     property public final Integer? topK;
     field public static final com.google.firebase.ai.OnDeviceConfig.Companion Companion;
diff --git a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/GenerativeModel.kt b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/GenerativeModel.kt
index 389d54434c5..b659958d33f 100644
--- a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/GenerativeModel.kt
+++ b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/GenerativeModel.kt
@@ -356,14 +356,31 @@ internal constructor(
     @OptIn(PublicPreviewAPI::class)
     internal fun buildOnDeviceModelProvider(
       modelOption: OnDeviceModelOption?
-    ): GenerativeModelProvider =
-      onDeviceFactoryProvider?.let {
+    ): GenerativeModelProvider {
+      // if developer didn't pass a OnDeviceConfig.SystemInstruction, we use GenerativeModel.systemInstruction
+      val effectiveOnDeviceConfig =
+        if (onDeviceConfig.systemInstruction == null && systemInstruction != null) {
+          OnDeviceConfig(
+            mode = onDeviceConfig.mode,
+            maxOutputTokens = onDeviceConfig.maxOutputTokens,
+            temperature = onDeviceConfig.temperature,
+            topK = onDeviceConfig.topK,
+            seed = onDeviceConfig.seed,
+            candidateCount = onDeviceConfig.candidateCount,
+            modelOption = onDeviceConfig.modelOption,
+            systemInstruction = systemInstruction
+          )
+        } else {
+          onDeviceConfig
+        }
+      return onDeviceFactoryProvider?.let {
         OnDeviceGenerativeModelProvider(
           it.newGenerativeModel(modelOption?.toInterop()),
-          onDeviceConfig
+          effectiveOnDeviceConfig
         )
       }
         ?: MissingOnDeviceGenerativeModelProvider()
+    }
 
     @PublicPreviewAPI
     internal fun getModelProvider(): GenerativeModelProvider =
diff --git a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/OnDeviceConfig.kt b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/OnDeviceConfig.kt
index 7b613ebaf8c..8860fa1cc76 100644
--- a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/OnDeviceConfig.kt
+++ b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/OnDeviceConfig.kt
@@ -16,6 +16,7 @@
 
 package com.google.firebase.ai
 
+import com.google.firebase.ai.type.Content
 import com.google.firebase.ai.type.FirebaseAIException
 import com.google.firebase.ai.type.PublicPreviewAPI
 
@@ -34,6 +35,7 @@ import com.google.firebase.ai.type.PublicPreviewAPI
  * @property candidateCount The number of generated responses to return. See [GenerationConfig] for
  * more detail. By default it's set to `1`.
  * @property modelOption Configuration for the on-device model selection and performance.
+ * @property systemInstruction Instructions that direct the model to behave in a certain way.
  */
 @PublicPreviewAPI
 public class OnDeviceConfig
@@ -45,7 +47,8 @@ constructor(
   public val topK: Int? = null,
   public val seed: Int? = null,
   public val candidateCount: Int = 1,
-  public val modelOption: OnDeviceModelOption? = null
+  public val modelOption: OnDeviceModelOption? = null,
+  public val systemInstruction: Content? = null
 ) {
 
   public companion object {
diff --git a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/generativemodel/OnDeviceGenerativeModelProvider.kt b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/generativemodel/OnDeviceGenerativeModelProvider.kt
index 5471db5dfdb..896dea968ff 100644
--- a/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/generativemodel/OnDeviceGenerativeModelProvider.kt
+++ b/ai-logic/firebase-ai/src/main/kotlin/com/google/firebase/ai/generativemodel/OnDeviceGenerativeModelProvider.kt
@@ -215,6 +215,13 @@ internal class OnDeviceGenerativeModelProvider(
             )
         }
         .firstOrNull()
+    val systemInstructionPart =
+      onDeviceConfig.systemInstruction
+        ?.parts
+        ?.filterIsInstance<TextPart>()
+        ?.joinToString("") { it.text }
+        ?.takeIf { it.isNotEmpty() }
+        ?.let { OnDeviceTextPart(it) }
     return OnDeviceGenerateContentRequest(
       text = OnDeviceTextPart(text),
       image = image?.let { OnDeviceImagePart(it.image) },
@@ -222,7 +229,8 @@ internal class OnDeviceGenerativeModelProvider(
       topK = onDeviceConfig.topK,
       seed = onDeviceConfig.seed,
       candidateCount = onDeviceConfig.candidateCount,
-      maxOutputTokens = onDeviceConfig.maxOutputTokens
+      maxOutputTokens = onDeviceConfig.maxOutputTokens,
+      systemInstruction = systemInstructionPart
     )
   }
 
diff --git a/ai-logic/firebase-ai/src/test/java/com/google/firebase/ai/generativeModel/OnDeviceGenerativeModelProviderTests.kt b/ai-logic/firebase-ai/src/test/java/com/google/firebase/ai/generativeModel/OnDeviceGenerativeModelProviderTests.kt
index 9ef22433578..dec63179753 100644
--- a/ai-logic/firebase-ai/src/test/java/com/google/firebase/ai/generativeModel/OnDeviceGenerativeModelProviderTests.kt
+++ b/ai-logic/firebase-ai/src/test/java/com/google/firebase/ai/generativeModel/OnDeviceGenerativeModelProviderTests.kt
@@ -24,6 +24,7 @@ import com.google.firebase.ai.ondevice.interop.Candidate as OnDeviceCandidate
 import com.google.firebase.ai.ondevice.interop.CountTokensResponse as OnDeviceCountTokensResponse
 import com.google.firebase.ai.ondevice.interop.FinishReason as OnDeviceFinishReason
 import com.google.firebase.ai.ondevice.interop.FirebaseAIOnDeviceNotAvailableException
+import com.google.firebase.ai.ondevice.interop.GenerateContentRequest as OnDeviceGenerateContentRequest
 import com.google.firebase.ai.ondevice.interop.GenerateContentResponse as OnDeviceGenerateContentResponse
 import com.google.firebase.ai.ondevice.interop.GenerativeModel as OnDeviceGenerativeModel
 import com.google.firebase.ai.toInterop
@@ -37,6 +38,7 @@ import io.kotest.matchers.shouldBe
 import io.mockk.coEvery
 import io.mockk.every
 import io.mockk.mockk
+import io.mockk.slot
 import kotlinx.coroutines.flow.first
 import kotlinx.coroutines.flow.flowOf
 import kotlinx.coroutines.runBlocking
@@ -153,4 +155,29 @@ internal class OnDeviceGenerativeModelProviderTests {
     interopConfig.modelConfig?.preference shouldBe
       com.google.firebase.ai.ondevice.interop.ModelPreference.FAST
   }
+
+  @Test
+  fun `generateContent passes systemInstruction in OnDeviceConfig to interop request`(): Unit =
+    runBlocking {
+      coEvery { onDeviceModel.isAvailable() } returns true
+      val systemInstruction = Content(parts = listOf(TextPart("system instruction rule")))
+      val configWithSystemInstruction =
+        OnDeviceConfig(
+          mode = InferenceMode.ONLY_ON_DEVICE,
+          temperature = 0.5f,
+          systemInstruction = systemInstruction
+        )
+      val providerWithSystemInstruction =
+        OnDeviceGenerativeModelProvider(onDeviceModel, configWithSystemInstruction)
+
+      val requestSlot = slot<OnDeviceGenerateContentRequest>()
+      coEvery { onDeviceModel.generateContent(capture(requestSlot)) } returns
+        OnDeviceGenerateContentResponse(
+          listOf(OnDeviceCandidate("generated text", OnDeviceFinishReason.STOP))
+        )
+
+      providerWithSystemInstruction.generateContent(prompt)
+
+      requestSlot.captured.systemInstruction?.text shouldBe "system instruction rule"
+    }
 }
diff --git a/gradle/libs.versions.toml b/gradle/libs.versions.toml
index 618f5ee8c6e..762d7b6a26e 100644
--- a/gradle/libs.versions.toml
+++ b/gradle/libs.versions.toml
@@ -28,7 +28,7 @@ firebaseAnnotations = "17.0.0"
 firebaseCommon = "22.0.1"
 firebaseComponents = "19.0.0"
 firebaseCrashlyticsGradle = "3.0.4"
-genaiPrompt = "1.0.0-beta2"
+genaiPrompt = "1.0.0-beta4"
 glide = "5.0.5"
 googleApiClient = "2.8.1"
 googleServices = "4.3.15"