@xybrid/react-native 0.10.1 → 0.13.0

package/README.md

@@ -215,8 +215,9 @@ const download = await loader.download();
 for await (const status of download!.progress()) setProgress(status.progress);
 const model = await loader.load();
 
-// Answer from the cloud while the weights download (needs an API key).
-const fast = await ModelLoader.fromRegistrySpeculative('qwen3.5-0.8b').load();
+// Answer from the cloud while the weights download (needs an API key and a
+// chat model the registry serves as a direct GGUF download).
+const fast = await ModelLoader.fromRegistrySpeculative('lfm2.5-350m').load();
 await fast.isCloudServing(); // true until the local weights land
 
 await Xybrid.modelCacheStatus();       // bytes and models on disk

package/android/build.gradle

@@ -57,5 +57,5 @@ dependencies {
 
   // The Kotlin SDK. Kept in lockstep with the workspace version by
   // tools/scripts/version-sync.sh (`just bump-version`).
-  implementation "ai.xybrid:xybrid-kotlin:0.10.1"
+  implementation "ai.xybrid:xybrid-kotlin:0.13.0"
 }

package/android/src/main/java/ai/xybrid/reactnative/XybridModule.kt

@@ -429,9 +429,12 @@ class XybridModule(reactContext: ReactApplicationContext) : NativeXybridSpec(rea
             handles.dispose(stream)
             mapOf("kind" to "complete", "result" to XybridCodec.encodeResult(result))
           }
+          XybridStreamEventKind.AUDIO -> throw BridgeException.InvalidArgument(
+            "Token streams cannot deliver audio; use run for batch speech"
+          )
         }
       } catch (t: Throwable) {
-        // A failed pull already closed the bolt session.
+        // Dispose also aborts a stream rejected for an unsupported event.
         handles.dispose(stream)
         throw cancellationAware(t, s.cancel)
       }

package/ios/XybridModuleImpl.swift

@@ -616,9 +616,11 @@ public final class XybridModuleImpl: NSObject {
           let result = try entry.model.streamResult(streamId: entry.streamId)
           handles.dispose(stream)
           return ["kind": "complete", "result": XybridCodec.encodeResult(result)] as JSObject
+        case .audio:
+          throw BridgeError.invalidArgument("Token streams cannot deliver audio; use run for batch speech")
         }
       } catch {
-        // A failed pull already closed the bolt session.
+        // Dispose also aborts a stream rejected for an unsupported event.
         handles.dispose(stream)
         throw Self.cancellationAware(error, entry.cancel)
       }

package/ios/XybridSwift/Xybrid.swift

@@ -162,8 +162,9 @@ public enum Xybrid {
     /// ``ModelLoader/fromRegistrySpeculative(_:)`` opts in explicitly and is
     /// unaffected by this toggle. Off by default. Either way, speculation also
     /// needs a resolvable API key and a model that is not already cached —
-    /// ``ModelLoader/willSpeculate`` reports the combined answer for a
-    /// specific loader.
+    /// ``ModelLoader/willSpeculate`` checks both for a specific loader — and
+    /// the load asks the registry first, so only chat models are served this
+    /// way.
     public static func setSpeculativeCloud(_ enabled: Bool) {
         boltSetSpeculativeCloud(enabled)
     }
@@ -366,9 +367,10 @@ public struct ModelLoader: Sendable {
     ///
     /// ``load()`` returns almost immediately with a cloud-backed model that
     /// switches to on-device by itself once the download lands. Requires an API
-    /// key and an uncached model — otherwise it behaves exactly like
-    /// ``fromRegistry(_:)``, which ``willSpeculate`` reports up front. LLM/chat
-    /// models only.
+    /// key, an uncached model, and a chat model the registry serves as a direct
+    /// download (GGUF or vision-language, not a `.xyb` bundle) — otherwise it
+    /// behaves exactly like ``fromRegistry(_:)``. ``willSpeculate`` checks the
+    /// first two up front.
     ///
     /// Track the handover with ``XybridModel/isCloudServing()``,
     /// ``XybridModel/downloadStatus()`` and ``XybridModel/awaitDownload(timeoutMs:)``.
@@ -376,10 +378,13 @@ public struct ModelLoader: Sendable {
         Self(source: .registrySpeculative(id))
     }
 
-    /// Whether ``load()`` would actually speculate: speculation is possible for
-    /// this source, an API key resolves, and the model is not already cached.
+    /// Whether ``load()`` may speculate: speculation is possible for this
+    /// source, an API key resolves, and the model is not already cached.
     ///
-    /// Always `false` for non-speculative sources. Never touches the network.
+    /// Always `false` for non-speculative sources. Never touches the network,
+    /// so it cannot know the model type: ``load()`` asks the registry and
+    /// speculates only for a chat model. ``XybridModel/isCloudServing()``
+    /// reports the outcome.
     public var willSpeculate: Bool {
         guard case .registrySpeculative(let id) = source else { return false }
         return willSpeculateForModel(modelId: id)
@@ -643,6 +648,11 @@ private final class XybridTokenStreamState: @unchecked Sendable {
         case .complete:
             close()
             return nil
+        case .audio:
+            close()
+            throw XybridError.inferenceError(
+                message: "Token streams cannot deliver audio; use run for batch speech"
+            )
         }
     }
 

package/ios/XybridSwift/xybrid_bolt.swift

@@ -592,6 +592,42 @@ public struct XybridStreamToken: Hashable, Equatable, Sendable {
     }
 }
 
+/// Owned PCM16 LE packet. Offsets count samples per channel.
+public struct XybridTtsAudioChunk: Hashable, Equatable, Sendable {
+    public var pcm: Data
+    public var sampleRate: UInt32
+    public var channels: UInt32
+    public var firstSample: UInt64
+
+    public init(
+        pcm: Data,
+        sampleRate: UInt32,
+        channels: UInt32,
+        firstSample: UInt64
+    ) {
+        self.pcm = pcm
+        self.sampleRate = sampleRate
+        self.channels = channels
+        self.firstSample = firstSample
+    }
+
+    @inlinable static func decode(from reader: inout WireReader) -> XybridTtsAudioChunk {
+        XybridTtsAudioChunk(
+            pcm: reader.readBytes(),
+            sampleRate: reader.readU32(),
+            channels: reader.readU32(),
+            firstSample: reader.readU64()
+        )
+    }
+
+    @inlinable func encode(to writer: inout WireWriter) {
+        writer.writeBytes(self.pcm)
+        writer.writeU32(self.sampleRate)
+        writer.writeU32(self.channels)
+        writer.writeU64(self.firstSample)
+    }
+}
+
 /// One pull from a streaming inference session.
 ///
 /// This is a flat record instead of a data-carrying enum because the pinned
@@ -603,19 +639,22 @@ public struct XybridStreamToken: Hashable, Equatable, Sendable {
 public struct XybridStreamEvent: Hashable, Equatable, Sendable {
     public var kind: XybridStreamEventKind
     public var token: XybridStreamToken?
+    public var audio: XybridTtsAudioChunk?
 
-    public init(kind: XybridStreamEventKind, token: XybridStreamToken?) {
+    public init(kind: XybridStreamEventKind, token: XybridStreamToken?, audio: XybridTtsAudioChunk?) {
         self.kind = kind
         self.token = token
+        self.audio = audio
     }
 
     @inlinable static func decode(from reader: inout WireReader) -> XybridStreamEvent {
-        XybridStreamEvent(kind: XybridStreamEventKind(rawValue: reader.readI32())!, token: reader.readOptional { reader in XybridStreamToken.decode(from: &reader) })
+        XybridStreamEvent(kind: XybridStreamEventKind(rawValue: reader.readI32())!, token: reader.readOptional { reader in XybridStreamToken.decode(from: &reader) }, audio: reader.readOptional { reader in XybridTtsAudioChunk.decode(from: &reader) })
     }
 
     @inlinable func encode(to writer: inout WireWriter) {
         writer.writeI32(self.kind.rawValue)
         writer.writeOptional(self.token) { writer, boltffiValue0 in boltffiValue0.encode(to: &writer) }
+        writer.writeOptional(self.audio) { writer, boltffiValue0 in boltffiValue0.encode(to: &writer) }
     }
 }
 
@@ -1122,6 +1161,7 @@ public enum XybridDownloadState: Int32, Hashable, Sendable, CaseIterable {
 public enum XybridStreamEventKind: Int32, Hashable, Sendable, CaseIterable {
     case token = 0
     case complete = 1
+    case audio = 2
 
     @usableFromInline init(fromC c: Int32) {
         self = XybridStreamEventKind(rawValue: c)!
@@ -1414,9 +1454,11 @@ public final class XybridModel {
     /// download in the background.
     ///
     /// Returns almost immediately instead of blocking on the download. Requires
-    /// a resolvable API key and an uncached model; otherwise it behaves exactly
-    /// like `from_registry`. Poll `download_status` for progress and
-    /// `is_cloud_serving` to know which leg is answering. LLM/chat models only.
+    /// a resolvable API key, an uncached model, and a chat model the registry
+    /// serves as a direct download (GGUF or vision-language, not a `.xyb`
+    /// bundle); otherwise it behaves exactly like `from_registry`. Poll
+    /// `download_status` for progress and `is_cloud_serving` to know which leg
+    /// is answering.
     public init(fromRegistrySpeculative id: String) throws {
         let boltffiIdBytes = boltffiEncode { boltffiIdWriter in boltffiIdWriter.writeString(id) }
         let boltffiHandle = try boltffiIdBytes.withUnsafeBufferPointer { boltffiIdBuffer in
@@ -2428,11 +2470,12 @@ public func setPlatformUrl(url: String) {
     }
 }
 
-/// Enable speculative cloud fallback globally: a registry model that isn't
-/// downloaded yet is served from the gateway while the weights download.
+/// Enable speculative cloud fallback globally: a registry chat model that
+/// isn't downloaded yet is served from the gateway while the weights download.
 ///
-/// LLM/chat only — prefer `XybridModel.fromRegistrySpeculative` when the app
-/// also loads ASR/TTS models, which cannot be served this way.
+/// An uncached load asks the registry first, so other models (ASR, TTS,
+/// `.xyb` bundles) download and load as usual.
+/// `XybridModel.fromRegistrySpeculative` opts a single load in instead.
 public func setSpeculativeCloud(enabled: Bool) {
     boltffi_function_xybrid_bolt_set_speculative_cloud(enabled)
 }
@@ -2447,11 +2490,13 @@ public func isSpeculativeCloudEnabled() -> Bool {
     return boltffi_function_xybrid_bolt_is_speculative_cloud_enabled()
 }
 
-/// Whether `XybridModel::from_registry_speculative(model_id)` would actually
-/// speculate: an API key resolves and the model is not already cached.
+/// Whether `XybridModel::from_registry_speculative(model_id)` may speculate:
+/// an API key resolves and the model is not already cached.
 ///
-/// Lets the hand-written Swift/Kotlin loader facades answer "will this
-/// speculate?" before loading. Never touches the network.
+/// Lets the hand-written Swift/Kotlin loader facades answer "can this
+/// speculate?" before loading. Never touches the network, so it cannot know
+/// the model type: the load asks the registry and speculates only for a chat
+/// model.
 public func willSpeculateForModel(modelId: String) -> Bool {
     let boltffiModelIdBytes = boltffiEncode { boltffiModelIdWriter in boltffiModelIdWriter.writeString(modelId) }
     return boltffiModelIdBytes.withUnsafeBufferPointer { boltffiModelIdBuffer in

package/lib/commonjs/model.js

@@ -45,8 +45,10 @@ class ModelLoader {
      * Serve from the cloud gateway while the registry weights download in the
      * background, instead of blocking on the download. `load()` then resolves
      * almost immediately with a model that switches to on-device by itself.
-     * Needs an API key and an uncached model — {@link willSpeculate} tells you
-     * up front — otherwise it behaves like {@link fromRegistry}. LLM/chat only.
+     * Needs an API key, an uncached model ({@link willSpeculate} checks both up
+     * front) and a chat model the registry serves as a direct download (GGUF or
+     * vision-language, not a `.xyb` bundle) — otherwise it behaves like
+     * {@link fromRegistry}.
      */
     static fromRegistrySpeculative(id) {
         return new ModelLoader({ kind: 'registrySpeculative', id });
@@ -72,8 +74,10 @@ class ModelLoader {
         return new Model(await NativeXybrid_1.default.loadModel(toWireSource(this.source)));
     }
     /**
-     * Whether {@link load} would actually speculate: a speculative source, an
-     * API key, and a model not yet cached. Never touches the network.
+     * Whether {@link load} may speculate: a speculative source, an API key, and
+     * a model not yet cached. Never touches the network, so it cannot know the
+     * model type: `load()` asks the registry and speculates only for a chat
+     * model. `isCloudServing()` on the loaded model reports the outcome.
      */
     async willSpeculate() {
         if (this.source.kind !== 'registrySpeculative')

package/lib/commonjs/xybrid.js

@@ -50,9 +50,9 @@ exports.Xybrid = {
     },
     /**
      * Default speculative cloud serving on or off for loads that don't opt in
-     * per load (`ModelLoader.fromRegistrySpeculative` always speculates). Needs
-     * an API key. LLM/chat only — prefer the per-load form when the app also
-     * loads ASR/TTS models.
+     * per load (`ModelLoader.fromRegistrySpeculative` opts in itself). Needs an
+     * API key. An uncached load asks the registry first, so only chat models
+     * are served this way; ASR, TTS and `.xyb` bundles load as usual.
      */
     setSpeculativeCloud(enabled) {
         return NativeXybrid_1.default.setSpeculativeCloud(enabled);

package/lib/module/model.js

@@ -39,8 +39,10 @@ export class ModelLoader {
      * Serve from the cloud gateway while the registry weights download in the
      * background, instead of blocking on the download. `load()` then resolves
      * almost immediately with a model that switches to on-device by itself.
-     * Needs an API key and an uncached model — {@link willSpeculate} tells you
-     * up front — otherwise it behaves like {@link fromRegistry}. LLM/chat only.
+     * Needs an API key, an uncached model ({@link willSpeculate} checks both up
+     * front) and a chat model the registry serves as a direct download (GGUF or
+     * vision-language, not a `.xyb` bundle) — otherwise it behaves like
+     * {@link fromRegistry}.
      */
     static fromRegistrySpeculative(id) {
         return new ModelLoader({ kind: 'registrySpeculative', id });
@@ -66,8 +68,10 @@ export class ModelLoader {
         return new Model(await NativeXybrid.loadModel(toWireSource(this.source)));
     }
     /**
-     * Whether {@link load} would actually speculate: a speculative source, an
-     * API key, and a model not yet cached. Never touches the network.
+     * Whether {@link load} may speculate: a speculative source, an API key, and
+     * a model not yet cached. Never touches the network, so it cannot know the
+     * model type: `load()` asks the registry and speculates only for a chat
+     * model. `isCloudServing()` on the loaded model reports the outcome.
      */
     async willSpeculate() {
         if (this.source.kind !== 'registrySpeculative')

package/lib/module/xybrid.js

@@ -44,9 +44,9 @@ export const Xybrid = {
     },
     /**
      * Default speculative cloud serving on or off for loads that don't opt in
-     * per load (`ModelLoader.fromRegistrySpeculative` always speculates). Needs
-     * an API key. LLM/chat only — prefer the per-load form when the app also
-     * loads ASR/TTS models.
+     * per load (`ModelLoader.fromRegistrySpeculative` opts in itself). Needs an
+     * API key. An uncached load asks the registry first, so only chat models
+     * are served this way; ASR, TTS and `.xyb` bundles load as usual.
      */
     setSpeculativeCloud(enabled) {
         return NativeXybrid.setSpeculativeCloud(enabled);

package/lib/typescript/model.d.ts

@@ -66,8 +66,10 @@ export declare class ModelLoader {
      * Serve from the cloud gateway while the registry weights download in the
      * background, instead of blocking on the download. `load()` then resolves
      * almost immediately with a model that switches to on-device by itself.
-     * Needs an API key and an uncached model — {@link willSpeculate} tells you
-     * up front — otherwise it behaves like {@link fromRegistry}. LLM/chat only.
+     * Needs an API key, an uncached model ({@link willSpeculate} checks both up
+     * front) and a chat model the registry serves as a direct download (GGUF or
+     * vision-language, not a `.xyb` bundle) — otherwise it behaves like
+     * {@link fromRegistry}.
      */
     static fromRegistrySpeculative(id: string): ModelLoader;
     /** A local `.xyb` bundle. */
@@ -83,8 +85,10 @@ export declare class ModelLoader {
     /** Resolve, download if needed, and load. Rejects with a typed `xybrid_*` code. */
     load(): Promise<Model>;
     /**
-     * Whether {@link load} would actually speculate: a speculative source, an
-     * API key, and a model not yet cached. Never touches the network.
+     * Whether {@link load} may speculate: a speculative source, an API key, and
+     * a model not yet cached. Never touches the network, so it cannot know the
+     * model type: `load()` asks the registry and speculates only for a chat
+     * model. `isCloudServing()` on the loaded model reports the outcome.
      */
     willSpeculate(): Promise<boolean>;
     /**

package/lib/typescript/model.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model.d.ts","sourceRoot":"","sources":["../../src/model.ts"],"names":[],"mappings":"AACA,OAAO,EAAc,KAAK,eAAe,EAAE,MAAM,UAAU,CAAC;AAC5D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,WAAW,CAAC;AACrD,OAAO,EAAE,gBAAgB,EAAE,MAAM,WAAW,CAAC;AAC7C,OAAO,KAAK,EACV,eAAe,EACf,cAAc,EACd,QAAQ,EACR,gBAAgB,EAChB,eAAe,EACf,SAAS,EACT,WAAW,EACX,eAAe,EACf,SAAS,EACV,MAAM,SAAS,CAAC;AAYjB;;;;GAIG;AACH,MAAM,WAAW,UAAU;IACzB,6CAA6C;IAC7C,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,sDAAsD;IACtD,OAAO,CAAC,EAAE,eAAe,EAAE,CAAC;IAC5B,6EAA6E;IAC7E,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,4DAA4D;IAC5D,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,4DAA4D;IAC5D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,OAAO,CAAC,EAAE,mBAAmB,CAAC;IAC9B;;;;OAIG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;CAC1B;AAED,kFAAkF;AAClF,MAAM,MAAM,WAAW,GACnB;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAChC;IAAE,IAAI,EAAE,qBAAqB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAC3C;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAChC;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GACnC;IAAE,IAAI,EAAE,aAAa,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,MAAM,CAAA;CAAE,GACxD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAoBxC;;;GAGG;AACH,qBAAa,WAAW;IACF,QAAQ,CAAC,MAAM,EAAE,WAAW;IAAhD,OAAO;IAEP,2BAA2B;IAC3B,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,WAAW,GAAG,WAAW;IAI7C,8DAA8D;IAC9D,MAAM,CAAC,YAAY,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW;IAI5C;;;;;;OAMG;IACH,MAAM,CAAC,uBAAuB,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW;IAIvD,6BAA6B;IAC7B,MAAM,CAAC,UAAU,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI5C,0DAA0D;IAC1D,MAAM,CAAC,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI/C,uFAAuF;IACvF,MAAM,CAAC,eAAe,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,WAAW;IAItF,gFAAgF;IAChF,MAAM,CAAC,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI/C,mFAAmF;IAC7E,IAAI,IAAI,OAAO,CAAC,KAAK,CAAC;IAI5B;;;OAGG;IACG,aAAa,IAAI,OAAO,CAAC,OAAO,CAAC;IAKvC;;;;;OAKG;IACG,QAAQ,CAAC,OAAO,GAAE;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,OAAO,CAAC,aAAa,GAAG,IAAI,CAAC;CAKnF;AAED,+EAA+E;AAC/E,qBAAa,KAAK;IAGd,iEAAiE;IACjE,QAAQ,CAAC,MAAM,EAAE,MAAM;IAHzB,8CAA8C;;IAE5C,iEAAiE;IACxD,MAAM,EAAE,MAAM;IAGzB,qDAAqD;IAC/C,IAAI,IAAI,OAAO,CAAC,SAAS,CAAC;IAWhC,wFAAwF;IACxF,QAAQ,IAAI,OAAO,CAAC,OAAO,CAAC;IAI5B;;;OAGG;IACG,GAAG,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,GAAE,UAAe,GAAG,OAAO,CAAC,eAAe,CAAC;IAejF;;;;;;;;;;;;;;;;;OAiBG;IACI,YAAY,CACjB,QAAQ,EAAE,QAAQ,EAClB,OAAO,GAAE,UAAe,GACvB,cAAc,CAAC,WAAW,EAAE,eAAe,GAAG,SAAS,EAAE,IAAI,CAAC;IAiCjE;;;;OAIG;IACG,MAAM,CAAC,MAAM,GAAE,eAAoB,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAIrE,iFAAiF;IACjF,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;OAGG;IACH,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;;OAIG;IACH,cAAc,IAAI,OAAO,CAAC,OAAO,CAAC;IAIlC,gFAAgF;IAC1E,cAAc,IAAI,OAAO,CAAC,cAAc,CAAC;IAM/C;;;OAGG;IACH,gBAAgB,CAAC,OAAO,GAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,cAAc,CAAC,cAAc,EAAE,IAAI,EAAE,IAAI,CAAC;IAInG;;;OAGG;IACG,aAAa,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,cAAc,CAAC;IAM/D,mDAAmD;IAC7C,MAAM,IAAI,OAAO,CAAC,SAAS,EAAE,CAAC;IAIpC,sDAAsD;IAChD,YAAY,IAAI,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC;IAI/C,2BAA2B;IACrB,KAAK,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC;IAIvD;;;;OAIG;IACH,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAGzB;AA+BD;;;;GAIG;AACH,qBAAa,aAAa;IAEZ,QAAQ,CAAC,MAAM,EAAE,MAAM;IADnC,gBAAgB;gBACK,MAAM,EAAE,MAAM;IAEnC,wDAAwD;IAClD,MAAM,IAAI,OAAO,CAAC,cAAc,CAAC;IAMvC,iEAAiE;IACjE,KAAK,IAAI,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC;IAI/B,iEAAiE;IACjE,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;;;;;OAOG;IACH,QAAQ,CAAC,OAAO,GAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,cAAc,CAAC,cAAc,EAAE,IAAI,EAAE,IAAI,CAAC;IAI3F,mEAAmE;IACnE,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAGzB"}
+{"version":3,"file":"model.d.ts","sourceRoot":"","sources":["../../src/model.ts"],"names":[],"mappings":"AACA,OAAO,EAAc,KAAK,eAAe,EAAE,MAAM,UAAU,CAAC;AAC5D,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,WAAW,CAAC;AACrD,OAAO,EAAE,gBAAgB,EAAE,MAAM,WAAW,CAAC;AAC7C,OAAO,KAAK,EACV,eAAe,EACf,cAAc,EACd,QAAQ,EACR,gBAAgB,EAChB,eAAe,EACf,SAAS,EACT,WAAW,EACX,eAAe,EACf,SAAS,EACV,MAAM,SAAS,CAAC;AAYjB;;;;GAIG;AACH,MAAM,WAAW,UAAU;IACzB,6CAA6C;IAC7C,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,sDAAsD;IACtD,OAAO,CAAC,EAAE,eAAe,EAAE,CAAC;IAC5B,6EAA6E;IAC7E,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,4DAA4D;IAC5D,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,4DAA4D;IAC5D,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB;;;OAGG;IACH,OAAO,CAAC,EAAE,mBAAmB,CAAC;IAC9B;;;;OAIG;IACH,MAAM,CAAC,EAAE,eAAe,CAAC;CAC1B;AAED,kFAAkF;AAClF,MAAM,MAAM,WAAW,GACnB;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAChC;IAAE,IAAI,EAAE,qBAAqB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAC3C;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAChC;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GACnC;IAAE,IAAI,EAAE,aAAa,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,QAAQ,CAAC,EAAE,MAAM,CAAA;CAAE,GACxD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAoBxC;;;GAGG;AACH,qBAAa,WAAW;IACF,QAAQ,CAAC,MAAM,EAAE,WAAW;IAAhD,OAAO;IAEP,2BAA2B;IAC3B,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,WAAW,GAAG,WAAW;IAI7C,8DAA8D;IAC9D,MAAM,CAAC,YAAY,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW;IAI5C;;;;;;;;OAQG;IACH,MAAM,CAAC,uBAAuB,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW;IAIvD,6BAA6B;IAC7B,MAAM,CAAC,UAAU,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI5C,0DAA0D;IAC1D,MAAM,CAAC,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI/C,uFAAuF;IACvF,MAAM,CAAC,eAAe,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,GAAE;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,WAAW;IAItF,gFAAgF;IAChF,MAAM,CAAC,aAAa,CAAC,IAAI,EAAE,MAAM,GAAG,WAAW;IAI/C,mFAAmF;IAC7E,IAAI,IAAI,OAAO,CAAC,KAAK,CAAC;IAI5B;;;;;OAKG;IACG,aAAa,IAAI,OAAO,CAAC,OAAO,CAAC;IAKvC;;;;;OAKG;IACG,QAAQ,CAAC,OAAO,GAAE;QAAE,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,OAAO,CAAC,aAAa,GAAG,IAAI,CAAC;CAKnF;AAED,+EAA+E;AAC/E,qBAAa,KAAK;IAGd,iEAAiE;IACjE,QAAQ,CAAC,MAAM,EAAE,MAAM;IAHzB,8CAA8C;;IAE5C,iEAAiE;IACxD,MAAM,EAAE,MAAM;IAGzB,qDAAqD;IAC/C,IAAI,IAAI,OAAO,CAAC,SAAS,CAAC;IAWhC,wFAAwF;IACxF,QAAQ,IAAI,OAAO,CAAC,OAAO,CAAC;IAI5B;;;OAGG;IACG,GAAG,CAAC,QAAQ,EAAE,QAAQ,EAAE,OAAO,GAAE,UAAe,GAAG,OAAO,CAAC,eAAe,CAAC;IAejF;;;;;;;;;;;;;;;;;OAiBG;IACI,YAAY,CACjB,QAAQ,EAAE,QAAQ,EAClB,OAAO,GAAE,UAAe,GACvB,cAAc,CAAC,WAAW,EAAE,eAAe,GAAG,SAAS,EAAE,IAAI,CAAC;IAiCjE;;;;OAIG;IACG,MAAM,CAAC,MAAM,GAAE,eAAoB,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAIrE,iFAAiF;IACjF,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;OAGG;IACH,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;;OAIG;IACH,cAAc,IAAI,OAAO,CAAC,OAAO,CAAC;IAIlC,gFAAgF;IAC1E,cAAc,IAAI,OAAO,CAAC,cAAc,CAAC;IAM/C;;;OAGG;IACH,gBAAgB,CAAC,OAAO,GAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,cAAc,CAAC,cAAc,EAAE,IAAI,EAAE,IAAI,CAAC;IAInG;;;OAGG;IACG,aAAa,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,cAAc,CAAC;IAM/D,mDAAmD;IAC7C,MAAM,IAAI,OAAO,CAAC,SAAS,EAAE,CAAC;IAIpC,sDAAsD;IAChD,YAAY,IAAI,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC;IAI/C,2BAA2B;IACrB,KAAK,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC;IAIvD;;;;OAIG;IACH,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAGzB;AA+BD;;;;GAIG;AACH,qBAAa,aAAa;IAEZ,QAAQ,CAAC,MAAM,EAAE,MAAM;IADnC,gBAAgB;gBACK,MAAM,EAAE,MAAM;IAEnC,wDAAwD;IAClD,MAAM,IAAI,OAAO,CAAC,cAAc,CAAC;IAMvC,iEAAiE;IACjE,KAAK,IAAI,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC;IAI/B,iEAAiE;IACjE,MAAM,IAAI,OAAO,CAAC,IAAI,CAAC;IAIvB;;;;;;;OAOG;IACH,QAAQ,CAAC,OAAO,GAAE;QAAE,UAAU,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,cAAc,CAAC,cAAc,EAAE,IAAI,EAAE,IAAI,CAAC;IAI3F,mEAAmE;IACnE,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAGzB"}

package/lib/typescript/xybrid.d.ts

@@ -46,9 +46,9 @@ export declare const Xybrid: {
     setPlatformUrl(url: string): Promise<void>;
     /**
      * Default speculative cloud serving on or off for loads that don't opt in
-     * per load (`ModelLoader.fromRegistrySpeculative` always speculates). Needs
-     * an API key. LLM/chat only — prefer the per-load form when the app also
-     * loads ASR/TTS models.
+     * per load (`ModelLoader.fromRegistrySpeculative` opts in itself). Needs an
+     * API key. An uncached load asks the registry first, so only chat models
+     * are served this way; ASR, TTS and `.xyb` bundles load as usual.
      */
     setSpeculativeCloud(enabled: boolean): Promise<void>;
     isSpeculativeCloudEnabled(): Promise<boolean>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xybrid/react-native",
-  "version": "0.10.1",
+  "version": "0.13.0",
   "description": "On-device (and optionally cloud) ML inference for React Native: LLM, ASR, TTS, embeddings and vision, backed by the Xybrid Rust SDK",
   "main": "lib/commonjs/index.js",
   "module": "lib/module/index.js",
@@ -70,7 +70,7 @@
     "typescript": "^5.9.3"
   },
   "xybrid": {
-    "iosXcframeworkSha256": "3b012f6be628845d86ef274824de7b1e8c23c1b551b6965d18efd98268ee8b14"
+    "iosXcframeworkSha256": "a3a6abdadeea791a4fcc3fc005746b9115fced514725a0faf657774a22652ad9"
   },
   "codegenConfig": {
     "name": "RNXybridSpec",

package/src/model.ts

@@ -101,8 +101,10 @@ export class ModelLoader {
    * Serve from the cloud gateway while the registry weights download in the
    * background, instead of blocking on the download. `load()` then resolves
    * almost immediately with a model that switches to on-device by itself.
-   * Needs an API key and an uncached model — {@link willSpeculate} tells you
-   * up front — otherwise it behaves like {@link fromRegistry}. LLM/chat only.
+   * Needs an API key, an uncached model ({@link willSpeculate} checks both up
+   * front) and a chat model the registry serves as a direct download (GGUF or
+   * vision-language, not a `.xyb` bundle) — otherwise it behaves like
+   * {@link fromRegistry}.
    */
   static fromRegistrySpeculative(id: string): ModelLoader {
     return new ModelLoader({ kind: 'registrySpeculative', id });
@@ -134,8 +136,10 @@ export class ModelLoader {
   }
 
   /**
-   * Whether {@link load} would actually speculate: a speculative source, an
-   * API key, and a model not yet cached. Never touches the network.
+   * Whether {@link load} may speculate: a speculative source, an API key, and
+   * a model not yet cached. Never touches the network, so it cannot know the
+   * model type: `load()` asks the registry and speculates only for a chat
+   * model. `isCloudServing()` on the loaded model reports the outcome.
    */
   async willSpeculate(): Promise<boolean> {
     if (this.source.kind !== 'registrySpeculative') return false;

package/src/xybrid.ts

@@ -75,9 +75,9 @@ export const Xybrid = {
 
   /**
    * Default speculative cloud serving on or off for loads that don't opt in
-   * per load (`ModelLoader.fromRegistrySpeculative` always speculates). Needs
-   * an API key. LLM/chat only — prefer the per-load form when the app also
-   * loads ASR/TTS models.
+   * per load (`ModelLoader.fromRegistrySpeculative` opts in itself). Needs an
+   * API key. An uncached load asks the registry first, so only chat models
+   * are served this way; ASR, TTS and `.xyb` bundles load as usual.
    */
   setSpeculativeCloud(enabled: boolean): Promise<void> {
     return NativeXybrid.setSpeculativeCloud(enabled);