0.10.0

Note

The LiteRT-LM dependency in this release can cause checkout failures for Git LFS users, even when LiteRT support is disabled. Version 0.11.0 removes that dependency and the LiteRT trait and LiteRTLanguageModel API while retaining the other improvements in this release.

What's Changed

• Fix LlamaLanguageModel build against current llama.swift by @james-333i in #193
• Fix swift-format lint failure in TranscriptTests by @mattt in #207
• Send images per message in Ollama chat requests by @james-333i in #199
• Accept stringified primitives in Generable conversions by @james-333i in #201
• Cancel the generation when a response stream's consumer stops early by @james-333i in #202
• Adopt the MLX Memory API by @james-333i in #204
• Ingest decoder prompts in batch-sized chunks by @james-333i in #196
• Raise the macOS CI job timeout to 20 minutes by @mattt in #208
• Fix three places where constrained JSON generation ignores the model by @asaptf in #180
• Enable Metal GPU offload for llama.cpp models by default by @james-333i in #194
• Fix constrained sampling ignoring the allowed-token set by @james-333i in #197
• Add Swift 6.3 to CI by @mattt in #209
• Add an assistant prefill option to LlamaLanguageModel by @james-333i in #198
• Remove calls to the deprecated llama_set_warmup by @james-333i in #205
• Deliver streamed tokens live instead of after generation completes by @james-333i in #195
• Drop the duplicated trailing user turn in Foundation Models sessions by @james-333i in #217
• Widen the empty-array probe to merged item-start tokens by @james-333i in #212
• Carry Gemini thought signatures on the transcript by @NoWaY233851 in #219
• Add vision support to LlamaLanguageModel via mtmd by @james-333i in #213
• Render the Gemma 4 chat format when template detection fails by @james-333i in #214
• Require llama.swift 2.10549.0 and free the lazy bitmap for rejected video input by @mattt in #222
• Reuse a per-session context across chat exchanges by @mattt in #223
• Support tool calling in MLXLanguageModel.streamResponse() by @noorbhatia in #181
• Finish a response stream only after its transcript entries are committed by @mattt in #224
• Serialize lazy loading in LlamaLanguageModel by @mattt in #226
• Add tool calling to LlamaLanguageModel, including streaming by @mattt in #227
• Bump actions/checkout from 7.0.0 to 7.0.1 in the actions group across 1 directory by @dependabot[bot] in #179
• Route think and keep_alive to the top level of Ollama chat requests by @james-333i in #200
• Preserve Gemma tool types and response whitespace by @mattt in #228
• Add adaptive thinking and effort options for Anthropic by @ActuallyTaylor in #211
• Resolve tool schema $refs before MLX chat template rendering by @Adi2K in #177
• Forward all generation options to FoundationModels by @mattt in #230
• Preserve object and array shapes in union schemas by @mattt in #232
• Expose MLX sampler parameters and honor GenerationOptions.sampling by @mattt in #231
• Add LiteRTLanguageModel using the official LiteRT-LM Swift package by @mattt in #229
• Bump README install snippets to 0.10.0 by @mattt in #233

New Contributors

• @james-333i made their first contribution in #193
• @asaptf made their first contribution in #180
• @NoWaY233851 made their first contribution in #219
• @Adi2K made their first contribution in #177

Full Changelog: 0.9.0...0.10.0