Add LiteRTLanguageModel using the official LiteRT-LM Swift package

.github/workflows/ci.yml

@@ -10,6 +10,11 @@ permissions:
   contents: read
   pull-requests: write
 
+# LiteRT-LM's Swift package uses release XCFrameworks;
+# its Git LFS assets are not needed by SwiftPM.
+env:
+  GIT_LFS_SKIP_SMUDGE: "1"
+
 jobs:
   test-macos:
     name: Swift ${{ matrix.swift }} on macOS ${{ matrix.macos }} with Xcode ${{ matrix.xcode }}${{ matrix.traits != '' && format(' and --traits {0}', matrix.traits) || '' }}
@@ -63,10 +68,10 @@ jobs:
         run: swift format lint --strict --recursive .
 
       - name: Build
-        run: swift build --build-tests --traits MLX,Llama,CoreML${{ matrix.traits != '' && format(',{0}', matrix.traits) || '' }}
+        run: swift build --build-tests --traits MLX,Llama,CoreML,LiteRT${{ matrix.traits != '' && format(',{0}', matrix.traits) || '' }}
 
       - name: Test
-        run: swift test --skip-build --traits MLX,Llama,CoreML${{ matrix.traits != '' && format(',{0}', matrix.traits) || '' }}
+        run: swift test --skip-build --traits MLX,Llama,CoreML,LiteRT${{ matrix.traits != '' && format(',{0}', matrix.traits) || '' }}
 
   test-linux:
     name: Swift ${{ matrix.swift }} on Linux${{ matrix.traits != '' && format(' and --traits {0}', matrix.traits) || '' }}

Package.swift

@@ -25,6 +25,7 @@ let package = Package(
         .trait(name: "CoreML"),
         .trait(name: "MLX"),
         .trait(name: "Llama"),
+        .trait(name: "LiteRT"),
         .trait(name: "AsyncHTTPClient"),
         .default(enabledTraits: []),
     ],
@@ -43,6 +44,7 @@ let package = Package(
         .package(url: "https://github.com/mattt/llama.swift", .upToNextMajor(from: "2.10549.0")),
         .package(url: "https://github.com/mattt/PartialJSONDecoder", from: "1.0.0"),
         .package(url: "https://github.com/ml-explore/mlx-swift-lm", from: "3.31.4"),
+        .package(url: "https://github.com/google-ai-edge/LiteRT-LM", from: "0.17.0"),
         .package(url: "https://github.com/swiftlang/swift-syntax", from: "602.0.0"),
         .package(url: "https://github.com/swift-server/async-http-client.git", from: "1.24.0"),
     ],
@@ -77,7 +79,7 @@ let package = Package(
                 .product(
                     name: "HuggingFace",
                     package: "swift-huggingface",
-                    condition: .when(traits: ["MLX"])
+                    condition: .when(traits: ["MLX", "LiteRT"])
                 ),
                 .product(
                     name: "Tokenizers",
@@ -94,6 +96,11 @@ let package = Package(
                     package: "llama.swift",
                     condition: .when(traits: ["Llama"])
                 ),
+                .product(
+                    name: "LiteRTLM",
+                    package: "LiteRT-LM",
+                    condition: .when(platforms: [.iOS, .macOS], traits: ["LiteRT"])
+                ),
                 .product(
                     name: "AsyncHTTPClient",
                     package: "async-http-client",

README.md

@@ -73,6 +73,8 @@ session.toolExecutionDelegate = ToolExecutionObserver()
 - [x] [Core ML](https://developer.apple.com/documentation/coreml) models
 - [x] [MLX](https://github.com/ml-explore/mlx-swift) models
 - [x] [llama.cpp](https://github.com/ggml-org/llama.cpp) (GGUF models)
+- [x] [LiteRT-LM](https://github.com/google-ai-edge/litert-lm)
+  (`.litertlm` models, using the official Swift package)
 - [x] Ollama [HTTP API](https://github.com/ollama/ollama/blob/main/docs/api.md)
 - [x] Anthropic [Messages API](https://docs.claude.com/en/api/messages)
 - [x] Google [Gemini API](https://ai.google.dev/api/generate-content)
@@ -117,6 +119,9 @@ This results in smaller binary sizes and faster build times.
   (depends on `ml-explore/mlx-swift-lm`)
 - `Llama`: Enables llama.cpp support
   (requires `mattt/llama.swift`)
+- `LiteRT`: Enables LiteRT-LM support for Gemma 4 and other `.litertlm` models
+  (requires the official `google-ai-edge/LiteRT-LM` Swift package;
+  iOS and macOS only)
 
 By default, no traits are enabled.
 To enable specific traits, specify them in your package's dependencies:
@@ -366,13 +371,14 @@ Image support varies by provider:
 | Core ML                 | —               |
 | MLX                     | model-dependent |
 | llama.cpp               | —               |
+| LiteRT-LM               | model-dependent |
 | Ollama                  | model-dependent |
 | OpenAI                  | yes             |
 | Open Responses          | yes             |
 | Anthropic               | yes             |
 | Google Gemini           | yes             |
 
-For MLX and Ollama,
+For MLX, LiteRT-LM, and Ollama,
 use a vision-capable model 
 (for example, a VLM or `-vl` variant).
 
@@ -610,6 +616,39 @@ let response = try await session.respond(
 )
 ```
 
+### LiteRT-LM
+
+Runs `.litertlm` models (for example, Gemma 4) fully on-device
+via Google's [LiteRT-LM](https://github.com/google-ai-edge/litert-lm) runtime
+with Metal GPU acceleration
+(requires `LiteRT` trait;
+iOS and macOS only):
+
+```swift
+let model = LiteRTLanguageModel(modelFileURL: modelURL)
+
+let session = LanguageModelSession(model: model)
+let response = try await session.respond(to: "What is the capital of France?")
+```
+
+You can also load a `.litertlm` file from Hugging Face.
+The file is downloaded on first use
+using the Hub client's cache and authentication:
+
+```swift
+// Any Hugging Face repo
+let model = LiteRTLanguageModel(
+    huggingFaceRepo: "litert-community/gemma-4-E4B-it-litert-lm",
+    fileName: "gemma-4-E4B-it.litertlm")
+```
+
+Image inputs are supported for models that ship a vision tower
+(pass `visionBackend: .cpu()`).
+Structured generation is prompt-driven —
+the JSON schema is included in the prompt and the response is parsed.
+Tool calling is supported for `respond`
+(not yet for streaming).
+
 ### Ollama
 
 Run models locally via Ollama's