react-native-nobodywho 2.4.0 → 2.5.0

package/generated/cpp/nobodywho.cpp

@@ -119,7 +119,7 @@ void uniffi_nobodywho_uniffi_fn_free_rustchat(
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_constructor_rustchat_new(
     /*handle*/ uint64_t model, RustBuffer system_prompt, uint32_t context_size,
     RustBuffer template_variables, RustBuffer tools, RustBuffer sampler,
-    RustCallStatus *uniffi_out_err);
+    RustBuffer mtp, RustBuffer thread_count, RustCallStatus *uniffi_out_err);
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_method_rustchat_ask(
     /*handle*/ uint64_t ptr, RustBuffer message,
     RustCallStatus *uniffi_out_err);
@@ -141,6 +141,9 @@ uniffi_nobodywho_uniffi_fn_method_rustchat_get_system_prompt(
 /*handle*/ uint64_t
 uniffi_nobodywho_uniffi_fn_method_rustchat_get_template_variables(
     /*handle*/ uint64_t ptr);
+/*handle*/ uint64_t
+uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+    /*handle*/ uint64_t ptr);
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_method_rustchat_reset_context(
     /*handle*/ uint64_t ptr, RustBuffer system_prompt, RustBuffer tools);
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_method_rustchat_reset_history(
@@ -247,7 +250,8 @@ void uniffi_nobodywho_uniffi_fn_free_rusttts(
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_constructor_rusttts_new(
     RustBuffer source, RustBuffer architecture, RustBuffer voice,
     RustBuffer language, RustBuffer speed, RustBuffer steps,
-    RustBuffer silence_duration, RustBuffer device,
+    RustBuffer silence_duration, RustBuffer precision, RustBuffer temperature,
+    RustBuffer huggingface_token, RustBuffer device,
     RustCallStatus *uniffi_out_err);
 RustBuffer uniffi_nobodywho_uniffi_fn_method_rusttts_synthesize(
     /*handle*/ uint64_t ptr, RustBuffer text, RustCallStatus *uniffi_out_err);
@@ -321,11 +325,12 @@ RustBuffer uniffi_nobodywho_uniffi_fn_func_get_cached_models(
     RustCallStatus *uniffi_out_err);
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_func_load_model(
     RustBuffer model_path, int8_t use_gpu, RustBuffer projection_model_path,
-    RustBuffer on_download_progress);
+    RustBuffer draft_model_path, RustBuffer on_download_progress);
 /*handle*/ uint64_t uniffi_nobodywho_uniffi_fn_func_load_tts(
     RustBuffer source, RustBuffer architecture, RustBuffer voice,
     RustBuffer language, RustBuffer speed, RustBuffer steps,
-    RustBuffer silence_duration, RustBuffer device);
+    RustBuffer silence_duration, RustBuffer precision, RustBuffer temperature,
+    RustBuffer huggingface_token, RustBuffer device);
 /*handle*/ uint64_t
 uniffi_nobodywho_uniffi_fn_func_sampler_preset_constrain_with_grammar(
     RustBuffer grammar, RustCallStatus *uniffi_out_err);
@@ -500,6 +505,7 @@ uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_get_stats();
 uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_get_system_prompt();
 uint16_t
 uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables();
+uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate();
 uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context();
 uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_history();
 uint16_t uniffi_nobodywho_uniffi_checksum_method_rustchat_set_chat_history();
@@ -2734,7 +2740,7 @@ NativeNobodywho::NativeNobodywho(
           rt,
           jsi::PropNameID::forAscii(
               rt, "ubrn_uniffi_nobodywho_uniffi_fn_constructor_rustchat_new"),
-          6,
+          8,
           [this](jsi::Runtime &rt, const jsi::Value &thisVal,
                  const jsi::Value *args, size_t count) -> jsi::Value {
             return this
@@ -2836,6 +2842,18 @@ NativeNobodywho::NativeNobodywho(
             ->cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_get_template_variables(
                 rt, thisVal, args, count);
       });
+  props["ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate"] =
+      jsi::Function::createFromHostFunction(
+          rt,
+          jsi::PropNameID::forAscii(rt, "ubrn_uniffi_nobodywho_uniffi_fn_"
+                                        "method_rustchat_mtp_acceptance_rate"),
+          1,
+          [this](jsi::Runtime &rt, const jsi::Value &thisVal,
+                 const jsi::Value *args, size_t count) -> jsi::Value {
+            return this
+                ->cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+                    rt, thisVal, args, count);
+          });
   props["ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_reset_context"] =
       jsi::Function::createFromHostFunction(
           rt,
@@ -3360,7 +3378,7 @@ NativeNobodywho::NativeNobodywho(
           rt,
           jsi::PropNameID::forAscii(
               rt, "ubrn_uniffi_nobodywho_uniffi_fn_constructor_rusttts_new"),
-          8,
+          11,
           [this](jsi::Runtime &rt, const jsi::Value &thisVal,
                  const jsi::Value *args, size_t count) -> jsi::Value {
             return this->cpp_uniffi_nobodywho_uniffi_fn_constructor_rusttts_new(
@@ -3708,7 +3726,7 @@ NativeNobodywho::NativeNobodywho(
           rt,
           jsi::PropNameID::forAscii(
               rt, "ubrn_uniffi_nobodywho_uniffi_fn_func_load_model"),
-          4,
+          5,
           [this](jsi::Runtime &rt, const jsi::Value &thisVal,
                  const jsi::Value *args, size_t count) -> jsi::Value {
             return this->cpp_uniffi_nobodywho_uniffi_fn_func_load_model(
@@ -3719,7 +3737,7 @@ NativeNobodywho::NativeNobodywho(
           rt,
           jsi::PropNameID::forAscii(
               rt, "ubrn_uniffi_nobodywho_uniffi_fn_func_load_tts"),
-          8,
+          11,
           [this](jsi::Runtime &rt, const jsi::Value &thisVal,
                  const jsi::Value *args, size_t count) -> jsi::Value {
             return this->cpp_uniffi_nobodywho_uniffi_fn_func_load_tts(
@@ -4681,6 +4699,18 @@ NativeNobodywho::NativeNobodywho(
             ->cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables(
                 rt, thisVal, args, count);
       });
+  props["ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_"
+        "rate"] = jsi::Function::createFromHostFunction(
+      rt,
+      jsi::PropNameID::forAscii(rt, "ubrn_uniffi_nobodywho_uniffi_checksum_"
+                                    "method_rustchat_mtp_acceptance_rate"),
+      0,
+      [this](jsi::Runtime &rt, const jsi::Value &thisVal,
+             const jsi::Value *args, size_t count) -> jsi::Value {
+        return this
+            ->cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate(
+                rt, thisVal, args, count);
+      });
   props["ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context"] =
       jsi::Function::createFromHostFunction(
           rt,
@@ -5735,6 +5765,8 @@ NativeNobodywho::cpp_uniffi_nobodywho_uniffi_fn_constructor_rustchat_new(
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[3]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[4]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[5]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[6]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[7]),
       &status);
   uniffi::nobodywho::Bridging<RustCallStatus>::copyIntoJs(
       rt, callInvoker, status, args[count - 1]);
@@ -5849,6 +5881,17 @@ jsi::Value NativeNobodywho::
   return uniffi_jsi::Bridging</*handle*/ uint64_t>::toJs(rt, callInvoker,
                                                          value);
 }
+jsi::Value NativeNobodywho::
+    cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+        jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
+        size_t count) {
+  auto value = uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+      uniffi_jsi::Bridging</*handle*/ uint64_t>::fromJs(rt, callInvoker,
+                                                        args[0]));
+
+  return uniffi_jsi::Bridging</*handle*/ uint64_t>::toJs(rt, callInvoker,
+                                                         value);
+}
 jsi::Value
 NativeNobodywho::cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_reset_context(
     jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
@@ -6496,6 +6539,10 @@ NativeNobodywho::cpp_uniffi_nobodywho_uniffi_fn_constructor_rusttts_new(
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[5]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[6]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[7]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[8]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[9]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker,
+                                                      args[10]),
       &status);
   uniffi::nobodywho::Bridging<RustCallStatus>::copyIntoJs(
       rt, callInvoker, status, args[count - 1]);
@@ -6957,8 +7004,9 @@ jsi::Value NativeNobodywho::cpp_uniffi_nobodywho_uniffi_fn_func_load_model(
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[0]),
       uniffi_jsi::Bridging<int8_t>::fromJs(rt, callInvoker, args[1]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[2]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[3]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker,
-                                                      args[3]));
+                                                      args[4]));
 
   return uniffi_jsi::Bridging</*handle*/ uint64_t>::toJs(rt, callInvoker,
                                                          value);
@@ -6974,8 +7022,11 @@ jsi::Value NativeNobodywho::cpp_uniffi_nobodywho_uniffi_fn_func_load_tts(
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[4]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[5]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[6]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[7]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[8]),
+      uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker, args[9]),
       uniffi::nobodywho::Bridging<RustBuffer>::fromJs(rt, callInvoker,
-                                                      args[7]));
+                                                      args[10]));
 
   return uniffi_jsi::Bridging</*handle*/ uint64_t>::toJs(rt, callInvoker,
                                                          value);
@@ -7885,6 +7936,15 @@ jsi::Value NativeNobodywho::
 
   return uniffi_jsi::Bridging<uint16_t>::toJs(rt, callInvoker, value);
 }
+jsi::Value NativeNobodywho::
+    cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate(
+        jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
+        size_t count) {
+  auto value =
+      uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate();
+
+  return uniffi_jsi::Bridging<uint16_t>::toJs(rt, callInvoker, value);
+}
 jsi::Value NativeNobodywho::
     cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context(
         jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,

package/generated/cpp/nobodywho.hpp

@@ -64,6 +64,9 @@ protected:
   cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_get_template_variables(
       jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
       size_t count);
+  jsi::Value cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+      jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
+      size_t count);
   jsi::Value cpp_uniffi_nobodywho_uniffi_fn_method_rustchat_reset_context(
       jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
       size_t count);
@@ -557,6 +560,10 @@ protected:
   cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables(
       jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
       size_t count);
+  jsi::Value
+  cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate(
+      jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
+      size_t count);
   jsi::Value cpp_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context(
       jsi::Runtime &rt, const jsi::Value &thisVal, const jsi::Value *args,
       size_t count);

package/generated/ts/nobodywho-ffi.ts

@@ -20,7 +20,7 @@ interface NativeModuleInterface {
     ubrn_uniffi_internal_fn_func_ffi__arraybuffer_to_string(buffer: Uint8Array, uniffi_out_err: UniffiRustCallStatus): string;
     ubrn_uniffi_nobodywho_uniffi_fn_clone_rustchat(handle: bigint, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_free_rustchat(handle: bigint, uniffi_out_err: UniffiRustCallStatus): void;
-    ubrn_uniffi_nobodywho_uniffi_fn_constructor_rustchat_new(model: bigint, systemPrompt: Uint8Array, contextSize: number, templateVariables: Uint8Array, tools: Uint8Array, sampler: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
+    ubrn_uniffi_nobodywho_uniffi_fn_constructor_rustchat_new(model: bigint, systemPrompt: Uint8Array, contextSize: number, templateVariables: Uint8Array, tools: Uint8Array, sampler: Uint8Array, mtp: Uint8Array, threadCount: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_ask(ptr: bigint, message: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_ask_with_json_prompt(ptr: bigint, json: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_ask_with_prompt(ptr: bigint, parts: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
@@ -29,6 +29,7 @@ interface NativeModuleInterface {
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_get_stats(ptr: bigint): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_get_system_prompt(ptr: bigint): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_get_template_variables(ptr: bigint): bigint;
+    ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(ptr: bigint): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_reset_context(ptr: bigint, systemPrompt: Uint8Array, tools: Uint8Array): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_reset_history(ptr: bigint): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_set_chat_history(ptr: bigint, messages: Uint8Array): bigint;
@@ -73,7 +74,7 @@ interface NativeModuleInterface {
     ubrn_uniffi_nobodywho_uniffi_fn_method_rusttool_resolve_pending_call(ptr: bigint, callId: Uint8Array, result: Uint8Array, uniffi_out_err: UniffiRustCallStatus): void;
     ubrn_uniffi_nobodywho_uniffi_fn_clone_rusttts(handle: bigint, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_free_rusttts(handle: bigint, uniffi_out_err: UniffiRustCallStatus): void;
-    ubrn_uniffi_nobodywho_uniffi_fn_constructor_rusttts_new(source: Uint8Array, architecture: Uint8Array, voice: Uint8Array, language: Uint8Array, speed: Uint8Array, steps: Uint8Array, silenceDuration: Uint8Array, device: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
+    ubrn_uniffi_nobodywho_uniffi_fn_constructor_rusttts_new(source: Uint8Array, architecture: Uint8Array, voice: Uint8Array, language: Uint8Array, speed: Uint8Array, steps: Uint8Array, silenceDuration: Uint8Array, precision: Uint8Array, temperature: Uint8Array, huggingfaceToken: Uint8Array, device: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rusttts_synthesize(ptr: bigint, text: Uint8Array, uniffi_out_err: UniffiRustCallStatus): Uint8Array;
     ubrn_uniffi_nobodywho_uniffi_fn_method_rusttts_synthesize_async(ptr: bigint, text: Uint8Array): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_clone_samplerbuilder(handle: bigint, uniffi_out_err: UniffiRustCallStatus): bigint;
@@ -102,8 +103,8 @@ interface NativeModuleInterface {
     ubrn_uniffi_nobodywho_uniffi_fn_func_cosine_similarity(a: Uint8Array, b: Uint8Array, uniffi_out_err: UniffiRustCallStatus): number;
     ubrn_uniffi_nobodywho_uniffi_fn_func_download_model(modelPath: Uint8Array, headers: Uint8Array, onDownloadProgress: Uint8Array): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_func_get_cached_models(uniffi_out_err: UniffiRustCallStatus): Uint8Array;
-    ubrn_uniffi_nobodywho_uniffi_fn_func_load_model(modelPath: Uint8Array, useGpu: number, projectionModelPath: Uint8Array, onDownloadProgress: Uint8Array): bigint;
-    ubrn_uniffi_nobodywho_uniffi_fn_func_load_tts(source: Uint8Array, architecture: Uint8Array, voice: Uint8Array, language: Uint8Array, speed: Uint8Array, steps: Uint8Array, silenceDuration: Uint8Array, device: Uint8Array): bigint;
+    ubrn_uniffi_nobodywho_uniffi_fn_func_load_model(modelPath: Uint8Array, useGpu: number, projectionModelPath: Uint8Array, draftModelPath: Uint8Array, onDownloadProgress: Uint8Array): bigint;
+    ubrn_uniffi_nobodywho_uniffi_fn_func_load_tts(source: Uint8Array, architecture: Uint8Array, voice: Uint8Array, language: Uint8Array, speed: Uint8Array, steps: Uint8Array, silenceDuration: Uint8Array, precision: Uint8Array, temperature: Uint8Array, huggingfaceToken: Uint8Array, device: Uint8Array): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_func_sampler_preset_constrain_with_grammar(grammar: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_func_sampler_preset_constrain_with_json_schema(schema: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
     ubrn_uniffi_nobodywho_uniffi_fn_func_sampler_preset_constrain_with_regex(pattern: Uint8Array, uniffi_out_err: UniffiRustCallStatus): bigint;
@@ -187,6 +188,7 @@ interface NativeModuleInterface {
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_stats(): number;
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_system_prompt(): number;
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables(): number;
+    ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate(): number;
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context(): number;
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_history(): number;
     ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_set_chat_history(): number;

package/generated/ts/nobodywho.ts

@@ -145,20 +145,28 @@ export function getCachedModels(): Array<CachedModel> /*throws*/ {
 /**
  * Load a GGUF model from a local path or remote URL.
  *
- * Accepts local filesystem paths, `hf://owner/repo/file.gguf` for HuggingFace downloads,
- * or `https://` URLs. Downloaded models are cached automatically.
+ * Accepts local filesystem paths, `hf://owner/repo/file.gguf`, `https://` URLs,
+ * or `auto` for memory-based selection. Downloaded models are cached automatically.
+ *
+ * # MTP speculative decoding
+ *
+ * Pass `draft_model_path` pointing to a compatible MTP heads gguf (e.g.
+ * `mtp-gemma-4-E2B-it.gguf` for Gemma-4-E2B) to enable MTP
+ * speculative decoding on chats built from this model. Whether MTP is
+ * actually used is a per-chat decision — pass it through
+ * `Chat`-level config on the wrapping binding.
  *
  * This is a free function instead of an async constructor because
  * uniffi-bindgen-react-native generates invalid JS (`async static` instead
  * of `static async`) for async constructors.
  */
-export async function loadModel(modelPath: string, useGpu: boolean, projectionModelPath: string | undefined, onDownloadProgress: RustDownloadProgressCallback | undefined, asyncOpts_?: { signal: AbortSignal }): Promise<RustModelInterface> /*throws*/ {
+export async function loadModel(modelPath: string, useGpu: boolean, projectionModelPath: string | undefined, draftModelPath: string | undefined, onDownloadProgress: RustDownloadProgressCallback | undefined, asyncOpts_?: { signal: AbortSignal }): Promise<RustModelInterface> /*throws*/ {
     const __stack = uniffiIsDebug ? new Error().stack : undefined;
     try {
         return await uniffiRustCallAsync(
             /*rustCaller:*/ uniffiCaller,
             /*rustFutureFunc:*/ () => {
-                return nativeModule().ubrn_uniffi_nobodywho_uniffi_fn_func_load_model(FfiConverterString.lower(modelPath),FfiConverterBool.lower(useGpu),FfiConverterOptionalString.lower(projectionModelPath),FfiConverterOptionalTypeRustDownloadProgressCallback.lower(onDownloadProgress)
+                return nativeModule().ubrn_uniffi_nobodywho_uniffi_fn_func_load_model(FfiConverterString.lower(modelPath),FfiConverterBool.lower(useGpu),FfiConverterOptionalString.lower(projectionModelPath),FfiConverterOptionalString.lower(draftModelPath),FfiConverterOptionalTypeRustDownloadProgressCallback.lower(onDownloadProgress)
                 );
             },
             /*pollFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_poll_u64,
@@ -180,13 +188,13 @@ export async function loadModel(modelPath: string, useGpu: boolean, projectionMo
 /**
  * Create a TTS synthesizer.
  */
-export async function loadTts(source: string, architecture: string | undefined, voice: string | undefined, language: string | undefined, speed: /*f32*/number | undefined, steps: /*u32*/number | undefined, silenceDuration: /*f32*/number | undefined, device: string | undefined, asyncOpts_?: { signal: AbortSignal }): Promise<RustTtsInterface> /*throws*/ {
+export async function loadTts(source: string, architecture: string | undefined, voice: string | undefined, language: string | undefined, speed: /*f32*/number | undefined, steps: /*u32*/number | undefined, silenceDuration: /*f32*/number | undefined, precision: string | undefined, temperature: /*f32*/number | undefined, huggingfaceToken: string | undefined, device: string | undefined, asyncOpts_?: { signal: AbortSignal }): Promise<RustTtsInterface> /*throws*/ {
     const __stack = uniffiIsDebug ? new Error().stack : undefined;
     try {
         return await uniffiRustCallAsync(
             /*rustCaller:*/ uniffiCaller,
             /*rustFutureFunc:*/ () => {
-                return nativeModule().ubrn_uniffi_nobodywho_uniffi_fn_func_load_tts(FfiConverterString.lower(source),FfiConverterOptionalString.lower(architecture),FfiConverterOptionalString.lower(voice),FfiConverterOptionalString.lower(language),FfiConverterOptionalFloat32.lower(speed),FfiConverterOptionalUInt32.lower(steps),FfiConverterOptionalFloat32.lower(silenceDuration),FfiConverterOptionalString.lower(device)
+                return nativeModule().ubrn_uniffi_nobodywho_uniffi_fn_func_load_tts(FfiConverterString.lower(source),FfiConverterOptionalString.lower(architecture),FfiConverterOptionalString.lower(voice),FfiConverterOptionalString.lower(language),FfiConverterOptionalFloat32.lower(speed),FfiConverterOptionalUInt32.lower(steps),FfiConverterOptionalFloat32.lower(silenceDuration),FfiConverterOptionalString.lower(precision),FfiConverterOptionalFloat32.lower(temperature),FfiConverterOptionalString.lower(huggingfaceToken),FfiConverterOptionalString.lower(device)
                 );
             },
             /*pollFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_poll_u64,
@@ -656,6 +664,78 @@ const FfiConverterTypeChatStats = (() => {
 })();
 
 
+/**
+ * Tuning for MTP speculative decoding. Passing one to `RustChat::new`
+ * enables MTP; `null` runs the solo decode path. Requires the model to
+ * have been loaded with a compatible `draft_model_path`.
+ */
+export type MtpConfig = {
+    /**
+     * Maximum draft tokens proposed per speculative step (llama.cpp `n_max`).
+     * Higher values draft more per decode; returns diminish past ~4–6.
+     */
+    kMax: /*u32*/number,
+    /**
+     * Minimum draft-token probability the drafter will propose (llama.cpp
+     * `p_min`). `0.0` accepts all proposals; raise it to skip low-confidence
+     * drafts.
+     */
+    pMin: /*f32*/number
+}
+
+/**
+ * Generated factory for {@link MtpConfig} record objects.
+ */
+export const MtpConfig = (() => {
+    const defaults = () => ({kMax: 3,pMin: 0.0
+    });
+    const create = (() => {
+        return uniffiCreateRecord<MtpConfig, ReturnType<typeof defaults>>(defaults);
+    })();
+    return Object.freeze({
+        /**
+         * Create a frozen instance of {@link MtpConfig}, with defaults specified
+         * in Rust, in the {@link nobodywho} crate.
+         */
+        create,
+
+        /**
+         * Create a frozen instance of {@link MtpConfig}, with defaults specified
+         * in Rust, in the {@link nobodywho} crate.
+         */
+        new: create,
+
+        /**
+         * Defaults specified in the {@link nobodywho} crate.
+         */
+        defaults: () => Object.freeze(defaults()) as Partial<MtpConfig>,
+
+    });
+})();
+
+const FfiConverterTypeMtpConfig = (() => {
+    type TypeName = MtpConfig;
+    class FFIConverter extends AbstractFfiConverterByteArray<TypeName> {
+        read(from: RustBuffer): TypeName {
+            return {
+                kMax: FfiConverterUInt32.read(from), 
+                pMin: FfiConverterFloat32.read(from)
+            };
+        }
+        write(value: TypeName, into: RustBuffer): void {
+            FfiConverterUInt32.write(value.kMax, into);
+            FfiConverterFloat32.write(value.pMin, into);
+        }
+        allocationSize(value: TypeName): number {
+            return FfiConverterUInt32.allocationSize(value.kMax) + 
+            FfiConverterFloat32.allocationSize(value.pMin);
+            
+        }
+    };
+    return new FFIConverter();
+})();
+
+
 /**
  * A pending tool call waiting for resolution from the language binding.
  */
@@ -1418,6 +1498,12 @@ export interface RustChatInterface {
      * Get all template variables.
      */
     getTemplateVariables(asyncOpts_?: { signal: AbortSignal })  /*throws*/: Promise<Map<string, boolean>>;
+    /**
+     * MTP draft acceptance rate for the most recent generation, in `[0.0, 1.0]`.
+     *
+     * Resets each generation. `null` when MTP is disabled or no drafts were proposed.
+     */
+    mtpAcceptanceRate(asyncOpts_?: { signal: AbortSignal })  /*throws*/: Promise</*f32*/number | undefined>;
     /**
      * Reset the chat context with a new system prompt and tools.
      */
@@ -1469,8 +1555,18 @@ export class RustChat extends UniffiAbstractObject implements RustChatInterface
     readonly [pointerLiteralSymbol]: UniffiHandle;
     /**
      * Create a new chat session.
+     *
+     * Pass an `mtp` config to enable MTP speculative decoding for this
+     * chat; `null` disables it. Requires the `RustModel` to have been
+     * loaded with a compatible `draft_model_path`; otherwise construction
+     * fails. Adds around 5% to VRAM usage.
+     *
+     * `thread_count` is the number of CPU threads used for inference; `null`
+     * detects the device's physical core count (performance cores only, on
+     * Apple silicon), since hyperthreads and efficiency cores make inference
+     * slower. Clamped to the CPU count.
      */
-    constructor(model: RustModelInterface, systemPrompt: string | undefined, contextSize: /*u32*/number, templateVariables: Map<string, boolean> | undefined, tools: Array<RustToolInterface> | undefined, sampler: SamplerConfigInterface | undefined) /*throws*/ {
+    constructor(model: RustModelInterface, systemPrompt: string | undefined, contextSize: /*u32*/number, templateVariables: Map<string, boolean> | undefined, tools: Array<RustToolInterface> | undefined, sampler: SamplerConfigInterface | undefined, mtp: MtpConfig | undefined, threadCount: /*u32*/number | undefined) /*throws*/ {
         super();
         const pointer =
             
@@ -1484,6 +1580,8 @@ export class RustChat extends UniffiAbstractObject implements RustChatInterface
         FfiConverterOptionalMapStringBool.lower(templateVariables),
         FfiConverterOptionalArrayTypeRustTool.lower(tools),
         FfiConverterOptionalTypeSamplerConfig.lower(sampler),
+        FfiConverterOptionalTypeMtpConfig.lower(mtp),
+        FfiConverterOptionalUInt32.lower(threadCount),
                 callStatus);
             },
             /*liftString:*/ FfiConverterString.lift,
@@ -1700,6 +1798,39 @@ async  getTemplateVariables(asyncOpts_?: { signal: AbortSignal }): Promise<Map<s
     }
     }
     
+    /**
+     * MTP draft acceptance rate for the most recent generation, in `[0.0, 1.0]`.
+     *
+     * Resets each generation. `null` when MTP is disabled or no drafts were proposed.
+     */
+async  mtpAcceptanceRate(asyncOpts_?: { signal: AbortSignal }): Promise</*f32*/number | undefined> /*throws*/ {
+    const __stack = uniffiIsDebug ? new Error().stack : undefined;
+    try {
+        return await uniffiRustCallAsync(
+            /*rustCaller:*/ uniffiCaller,
+            /*rustFutureFunc:*/ () => {
+                return nativeModule().ubrn_uniffi_nobodywho_uniffi_fn_method_rustchat_mtp_acceptance_rate(
+                    uniffiTypeRustChatObjectFactory.clonePointer(this)
+                    
+                );
+            },
+            /*pollFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_poll_rust_buffer,
+            /*cancelFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_cancel_rust_buffer,
+            /*completeFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_complete_rust_buffer,
+            /*freeFunc:*/ nativeModule().ubrn_ffi_nobodywho_uniffi_rust_future_free_rust_buffer,
+            /*liftFunc:*/ FfiConverterOptionalFloat32.lift.bind(FfiConverterOptionalFloat32),
+            /*liftString:*/ FfiConverterString.lift,
+            /*asyncOpts:*/ asyncOpts_,
+            /*errorHandler:*/ FfiConverterTypeNobodyWhoError.lift.bind(FfiConverterTypeNobodyWhoError)
+        );
+    } catch (__error: any) {
+        if (uniffiIsDebug && __error instanceof Error) {
+            __error.stack = __stack;
+        }
+        throw __error;
+    }
+    }
+    
     /**
      * Reset the chat context with a new system prompt and tools.
      */
@@ -3218,7 +3349,7 @@ export class RustTts extends UniffiAbstractObject implements RustTtsInterface {
     /**
      * Create a TTS synthesizer.
      */
-    constructor(source: string, architecture: string | undefined, voice: string | undefined, language: string | undefined, speed: /*f32*/number | undefined, steps: /*u32*/number | undefined, silenceDuration: /*f32*/number | undefined, device: string | undefined) /*throws*/ {
+    constructor(source: string, architecture: string | undefined, voice: string | undefined, language: string | undefined, speed: /*f32*/number | undefined, steps: /*u32*/number | undefined, silenceDuration: /*f32*/number | undefined, precision: string | undefined, temperature: /*f32*/number | undefined, huggingfaceToken: string | undefined, device: string | undefined) /*throws*/ {
         super();
         const pointer =
             
@@ -3233,6 +3364,9 @@ export class RustTts extends UniffiAbstractObject implements RustTtsInterface {
         FfiConverterOptionalFloat32.lower(speed),
         FfiConverterOptionalUInt32.lower(steps),
         FfiConverterOptionalFloat32.lower(silenceDuration),
+        FfiConverterOptionalString.lower(precision),
+        FfiConverterOptionalFloat32.lower(temperature),
+        FfiConverterOptionalString.lower(huggingfaceToken),
         FfiConverterOptionalString.lower(device),
                 callStatus);
             },
@@ -3883,6 +4017,10 @@ const FfiConverterOptionalFloat32 = new FfiConverterOptional(FfiConverterFloat32
 const FfiConverterOptionalInt32 = new FfiConverterOptional(FfiConverterInt32);
 
 
+// FfiConverter for MtpConfig | undefined
+const FfiConverterOptionalTypeMtpConfig = new FfiConverterOptional(FfiConverterTypeMtpConfig);
+
+
 // FfiConverter for PendingToolCall | undefined
 const FfiConverterOptionalTypePendingToolCall = new FfiConverterOptional(FfiConverterTypePendingToolCall);
 
@@ -3985,10 +4123,10 @@ function uniffiEnsureInitialized() {
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_get_cached_models() !== 12002) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_func_get_cached_models");
     }
-    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_model() !== 33587) {
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_model() !== 22964) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_func_load_model");
     }
-    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_tts() !== 61935) {
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_load_tts() !== 26587) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_func_load_tts");
     }
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_func_sampler_preset_constrain_with_grammar() !== 13698) {
@@ -4048,6 +4186,9 @@ function uniffiEnsureInitialized() {
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables() !== 19616) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_method_rustchat_get_template_variables");
     }
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate() !== 727) {
+        throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_method_rustchat_mtp_acceptance_rate");
+    }
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context() !== 47191) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_method_rustchat_reset_context");
     }
@@ -4168,7 +4309,7 @@ function uniffiEnsureInitialized() {
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_method_samplerconfig_to_json() !== 51798) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_method_samplerconfig_to_json");
     }
-    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rustchat_new() !== 24505) {
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rustchat_new() !== 2313) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_constructor_rustchat_new");
     }
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rustcrossencoder_new() !== 9022) {
@@ -4186,7 +4327,7 @@ function uniffiEnsureInitialized() {
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rusttool_new_async() !== 54521) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_constructor_rusttool_new_async");
     }
-    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rusttts_new() !== 24110) {
+    if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_rusttts_new() !== 34899) {
         throw new UniffiInternalError.ApiChecksumMismatch("uniffi_nobodywho_uniffi_checksum_constructor_rusttts_new");
     }
     if (nativeModule().ubrn_uniffi_nobodywho_uniffi_checksum_constructor_samplerbuilder_new() !== 50214) {
@@ -4213,6 +4354,7 @@ export default Object.freeze({
     FfiConverterTypeCachedModel,
     FfiConverterTypeChatStats,
     FfiConverterTypeMessage,
+    FfiConverterTypeMtpConfig,
     FfiConverterTypeNobodyWhoError,
     FfiConverterTypePendingToolCall,
     FfiConverterTypePromptPart,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "react-native-nobodywho",
-  "version": "2.4.0",
+  "version": "2.5.0",
   "description": "Run LLMs locally with offline inference — React Native bindings",
   "main": "src/wrapper.ts",
   "types": "src/wrapper.ts",

package/src/chat.ts

@@ -1,6 +1,7 @@
 import {
   RustChat,
   SamplerConfig,
+  MtpConfig,
   type ChatStats,
 } from "../generated/ts/nobodywho";
 import { Model } from "./model";
@@ -31,6 +32,12 @@ export class Chat {
   /** @internal */
   private readonly _inner: RustChat;
 
+  /**
+   * `threadCount` is the number of CPU threads used for inference. Omit it to detect the
+   * device's physical core count (performance cores only, on Apple silicon) — hyperthreads
+   * and efficiency cores make inference slower, not faster. Lower it to leave CPU headroom
+   * for the rest of the app.
+   */
   constructor(opts: {
     model: Model;
     systemPrompt?: string;
@@ -38,6 +45,8 @@ export class Chat {
     templateVariables?: Record<string, boolean>;
     tools?: Tool[];
     sampler?: SamplerConfig;
+    mtp?: Partial<MtpConfig>;
+    threadCount?: number;
   }) {
     this._inner = new RustChat(
       opts.model._inner,
@@ -46,6 +55,8 @@ export class Chat {
       opts.templateVariables ? new Map(Object.entries(opts.templateVariables)) : undefined,
       opts.tools?.map((t) => t._inner) ?? undefined,
       opts.sampler ?? undefined,
+      opts.mtp !== undefined ? MtpConfig.create(opts.mtp) : undefined,
+      opts.threadCount ?? undefined,
     );
   }
 
@@ -65,17 +76,21 @@ export class Chat {
     modelPath: string;
     useGpu?: boolean;
     projectionModelPath?: string;
+    draftModelPath?: string;
     systemPrompt?: string;
     contextSize?: number;
     templateVariables?: Record<string, boolean>;
     tools?: Tool[];
     sampler?: SamplerConfig;
+    mtp?: Partial<MtpConfig>;
+    threadCount?: number;
     onDownloadProgress?: (downloaded: number, total: number) => void;
   }): Promise<Chat> {
     const model = await Model.load({
       modelPath: opts.modelPath,
       useGpu: opts.useGpu,
       projectionModelPath: opts.projectionModelPath,
+      draftModelPath: opts.draftModelPath,
       onDownloadProgress: opts.onDownloadProgress,
     });
     return new Chat({ model, ...opts });
@@ -173,6 +188,10 @@ export class Chat {
     return this._inner.getStats();
   }
 
+  async mtpAcceptanceRate(): Promise<number | undefined> {
+    return this._inner.mtpAcceptanceRate();
+  }
+
   /**
    * Immediately free the underlying Rust resources (model context, KV cache, etc.).
    * After calling this, the Chat instance is no longer usable.

package/src/model.ts

@@ -36,15 +36,19 @@ export class Model {
   /**
    * Load a GGUF model from a local path or remote URL.
    *
-   * @param opts.modelPath - Path to the GGUF model file, or a `hf://owner/repo/file.gguf` / `https://` URL to download and cache automatically
+   * @param opts.modelPath - Local path, `hf://` or `https://` URL, or `auto` for memory-based model selection
    * @param opts.useGpu - Use GPU acceleration if available (default: true)
    * @param opts.projectionModelPath - Path to a multimodal projector file for vision models
+   * @param opts.draftModelPath - Optional MTP draft-heads gguf. Loading it lets `Chat`
+   *   instances opt into MTP speculative decoding via `new Chat({ ..., mtp: true })`.
+   *   Adds around 5% to VRAM usage.
    * @param opts.onDownloadProgress - Invoked with `(downloaded, total)` byte counts while a remote model is being downloaded. Throttled to ~10 Hz with a guaranteed final emit on completion. Not invoked for cached/local files.
    */
   static async load(opts: {
     modelPath: string;
     useGpu?: boolean;
     projectionModelPath?: string;
+    draftModelPath?: string;
     onDownloadProgress?: (downloaded: number, total: number) => void;
   }): Promise<Model> {
     const cb = opts.onDownloadProgress;
@@ -55,6 +59,7 @@ export class Model {
       opts.modelPath,
       opts.useGpu ?? true,
       opts.projectionModelPath,
+      opts.draftModelPath,
       progressCallback,
     );
     return new Model(inner);

package/src/tts.ts

@@ -4,7 +4,7 @@ import {
   type RustTtsInterface,
 } from "../generated/ts/nobodywho";
 
-export type TtsArchitecture = "kokoro" | "supertonic";
+export type TtsArchitecture = "kokoro" | "pocket-tts" | "supertonic";
 export type TtsDevice = "auto" | "cpu" | "cuda";
 
 export type TtsOptions = {
@@ -15,6 +15,9 @@ export type TtsOptions = {
   speed?: number;
   steps?: number;
   silenceDuration?: number;
+  precision?: "int8" | "fp32";
+  temperature?: number;
+  huggingfaceToken?: string;
   device?: TtsDevice;
 };
 
@@ -33,6 +36,9 @@ export class Tts {
       opts.speed,
       opts.steps,
       opts.silenceDuration,
+      opts.precision,
+      opts.temperature,
+      opts.huggingfaceToken,
       opts.device ?? "auto",
     );
   }
@@ -53,6 +59,9 @@ export class Tts {
       opts.speed,
       opts.steps,
       opts.silenceDuration,
+      opts.precision,
+      opts.temperature,
+      opts.huggingfaceToken,
       opts.device ?? "auto",
     );
     return Tts.fromInner(inner);