Add Qwen 3.5 GGUF support and configurable chat-template kwargs (thinking toggle)

app/src/main/java/com/dark/tool_neuron/activity/ModelLoadingActivity.kt

@@ -235,7 +235,7 @@ fun ModelLoadingScreen(
                     // Create and insert config based on provider type
                     val config = when (model.providerType) {
                         ProviderType.GGUF -> {
-                            val defaultSchema = GgufEngineSchema()
+                            val defaultSchema = GgufEngineSchema.defaultsForModel(model.modelName)
                             ModelConfig(
                                 modelId = model.id,
                                 modelLoadingParams = defaultSchema.toLoadingJson(),
@@ -1060,4 +1060,4 @@ sealed class InstallState {
     data object Installing : InstallState()
     data object Installed : InstallState()
     data class Error(val message: String) : InstallState()
-}
+}

app/src/main/java/com/dark/tool_neuron/engine/GGUFEngine.kt

@@ -2,6 +2,7 @@ package com.dark.tool_neuron.engine
 
 import android.app.ActivityManager
 import android.content.Context
+import android.util.Log
 import com.dark.tool_neuron.models.table_schema.Model
 import com.dark.tool_neuron.models.table_schema.ModelConfig
 import com.dark.tool_neuron.models.engine_schema.DeviceTier
@@ -66,6 +67,7 @@ class GGUFEngine {
             if (inference.chatTemplate.isNotEmpty()) {
                 nativeLib.nativeSetChatTemplate(inference.chatTemplate)
             }
+            setChatTemplateKwargs(inference.chatTemplateKwargs)
         }
 
         success
@@ -117,6 +119,7 @@ class GGUFEngine {
             if (inference.chatTemplate.isNotEmpty()) {
                 nativeLib.nativeSetChatTemplate(inference.chatTemplate)
             }
+            setChatTemplateKwargs(inference.chatTemplateKwargs)
         }
 
         success
@@ -231,6 +234,16 @@ class GGUFEngine {
         nativeLib.nativeStopGeneration()
     }
 
+    private fun setChatTemplateKwargs(kwargsJson: String) {
+        if (kwargsJson.isBlank()) return
+        try {
+            setChatTemplateKwargsMethod?.invoke(nativeLib, kwargsJson)
+        } catch (e: ReflectiveOperationException) {
+            Log.d(TAG, "Chat template kwargs not supported in this native lib version (expected for older versions): ${e.message}")
+            // Backward-compatible with native libs that don't expose chat template kwargs
+        }
+    }
+
     suspend fun unload() = withContext(Dispatchers.IO) {
         if (isLoaded) {
             nativeLib.nativeRelease()
@@ -414,6 +427,15 @@ class GGUFEngine {
     fun hasToolsEnabled(): Boolean = !currentToolsJson.isNullOrEmpty()
 
     companion object {
+        private const val TAG = "GGUFEngine"
+        private val setChatTemplateKwargsMethod by lazy {
+            GGUFNativeLib::class.java.methods.firstOrNull {
+                it.name == "nativeSetChatTemplateKwargs" &&
+                        it.parameterTypes.size == 1 &&
+                        it.parameterTypes[0] == String::class.java
+            }
+        }
+
         /**
          * Detect device tier based on available RAM
          */

app/src/main/java/com/dark/tool_neuron/models/engine_schema/GgufEngineSchema.kt

@@ -4,6 +4,7 @@ import android.app.ActivityManager
 import android.content.Context
 import kotlinx.serialization.Serializable
 import kotlinx.serialization.json.Json
+import java.util.Locale
 
 enum class DeviceTier {
     LOW_END,    // < 4GB RAM
@@ -72,6 +73,7 @@ data class GgufInferenceParams(
     val maxTokens: Int = 4096,
     val systemPrompt: String = "",
     val chatTemplate: String = "",
+    val chatTemplateKwargs: String = "",
     val toolsJson: String = ""  // JSON array of tool definitions
 )
 
@@ -85,6 +87,9 @@ data class GgufEngineSchema(
     fun toInferenceJson(): String = json.encodeToString(inferenceParams)
 
     companion object {
+        private const val QWEN_THINKING_DISABLED_KWARGS = """{"enable_thinking": false}"""
+        // Match common Qwen 3.5 naming styles used across local files and HF repos.
+        private val qwen35ModelMarkers = listOf("qwen3.5", "qwen-3.5", "qwen3_5")
         private val json = Json {
             ignoreUnknownKeys = true
             encodeDefaults = true
@@ -109,5 +114,20 @@ data class GgufEngineSchema(
 
             return GgufEngineSchema(loading, inference)
         }
+
+        fun defaultsForModel(modelName: String): GgufEngineSchema {
+            val base = GgufEngineSchema()
+            if (!isQwen35Model(modelName)) return base
+            return base.copy(
+                inferenceParams = base.inferenceParams.copy(
+                    chatTemplateKwargs = QWEN_THINKING_DISABLED_KWARGS
+                )
+            )
+        }
+
+        private fun isQwen35Model(modelName: String): Boolean {
+            val normalized = modelName.lowercase(Locale.ROOT)
+            return qwen35ModelMarkers.any { normalized.contains(it) }
+        }
     }
-}
+}

app/src/main/java/com/dark/tool_neuron/repo/ModelRepoDataStore.kt

@@ -40,6 +40,14 @@ class ModelRepositoryDataStore(private val context: Context) {
                 isEnabled = true,
                 category = ModelCategory.GENERAL
             ),
+            HFModelRepository(
+                id = "qwen3_5_4b_instruct",
+                name = "Qwen3.5 Instruct (4B)",
+                repoPath = "Qwen/Qwen3.5-4B-Instruct-GGUF",
+                modelType = ModelType.GGUF,
+                isEnabled = true,
+                category = ModelCategory.GENERAL
+            ),
             HFModelRepository(
                 id = "liquidai-lfm2-350m",
                 name = "LFM2 350M",
@@ -180,4 +188,4 @@ class ModelRepositoryDataStore(private val context: Context) {
             if (it.id == repo.id) repo else it
         })
     }
-}
+}

app/src/main/java/com/dark/tool_neuron/service/ModelDownloadService.kt

@@ -574,7 +574,7 @@ class ModelDownloadService : Service() {
             }
 
             ProviderType.GGUF -> {
-                val ggufSchema = GgufEngineSchema()
+                val ggufSchema = GgufEngineSchema.defaultsForModel(modelName)
                 ModelConfig(
                     modelId = checksum,
                     modelLoadingParams = ggufSchema.toLoadingJson(),
@@ -671,4 +671,4 @@ class ModelDownloadService : Service() {
         super.onDestroy()
         serviceScope.cancel()
     }
-}
+}

app/src/main/java/com/dark/tool_neuron/ui/screen/ModelConfigEditorScreen.kt

@@ -463,6 +463,14 @@ private fun GgufConfigEditor(viewModel: ModelConfigEditorViewModel) {
                 multiline = true,
                 minLines = 3
             )
+
+            TextField(
+                label = "Chat Template Kwargs (Optional JSON)",
+                value = ggufConfig.inferenceParams.chatTemplateKwargs,
+                onValueChange = { viewModel.updateGgufChatTemplateKwargs(it) },
+                multiline = true,
+                minLines = 2
+            )
         }
     }
 }
@@ -884,4 +892,4 @@ private fun SuccessMessage() {
             )
         }
     }
-}
+}

app/src/main/java/com/dark/tool_neuron/viewmodel/ModelConfigEditorViewModel.kt

@@ -247,6 +247,12 @@ class ModelConfigEditorViewModel @Inject constructor(
         }
     }
 
+    fun updateGgufChatTemplateKwargs(value: String) {
+        _ggufConfig.update {
+            it.copy(inferenceParams = it.inferenceParams.copy(chatTemplateKwargs = value))
+        }
+    }
+
     // ==================== Diffusion Config Updates ====================
 
     fun updateDiffusionEmbeddingSize(value: Int) {
@@ -328,4 +334,4 @@ class ModelConfigEditorViewModel @Inject constructor(
             it.copy(showDiffusionStride = value)
         }
     }
-}
+}