2.25.4

What's Changed

• update to mlx-swift 0.25.4

• Expose MLXEmbedders as a library product by @finnvoor in #160

• add:Idefics3 image model support by @shuhongwu in #162

• pick up qwen2-vl fixes by @davidkoski in #163

• turn on docs build by @davidkoski in #166

• Fix SD generation in batch mode by @louen in #170

• Add VLMEval Sample Project by @rudrankriyam in #169

• fix llm-tool example resize bug by @JCSooHwanCho in #171

• llm-tool: fix prompt from file by @pcuenca in #176

• Improve handling of dependencies by @DePasqualeOrg in #185

• Revert conditional dependency by @DePasqualeOrg in #186

• correct handling of no-image case for qwen2-vl by @davidkoski in #179

• Update README.md by @davidkoski in #180

• Update LLMModelFactory.swift by @hasnat in #183

• Update README.md by @tzxdtc in #194

• Add video support for Qwen2-VL by @nvnsho in #187

• Support tool use and add example by @DePasqualeOrg in #174

• Use chat templates for vision models by @DePasqualeOrg in #173

• Fix visionOS support on the LVMEval Target by @ynagatomo in #199

• Fix messages format for text-only chats by @DePasqualeOrg in #207

• fix for #214 -- Qwen 2.5 fails to load: Key weight not found in Linear by @davidkoski in #215

• Support Phi-4-mini by @DePasqualeOrg in #216

• provide public accessors for ModelConfigurations from registries by @toffaletti in #219

• Make LLMModelFactory and VLMModelFactory inits public by @ibrahimcetin in #226

• Fix #218 - unable to load OpenELM by @davidkoski in #228

• Add argument for extra EOS token to llm-tool by @DePasqualeOrg in #217

• fix #231 -- use Embedding.asLinear by @davidkoski in #232

• Move common Registry types to MLXLMCommon by @ibrahimcetin in #229

• Add contains(id:) method to AbstractModelRegistry and ModelFactory by @ibrahimcetin in #233

• Improve errors by @DePasqualeOrg in #237

• Make initializer and some methods public in GenerateResult and GenerateParameters by @Alessan-git in #239

• Equatable conformance for ModelConfiguration and ModelConfiguration.Identifier by @Alessan-git in #246

• positionIds are not weights by @davidkoski in #251

• Improved token generation: single-token version, AsyncStream functionality, and supporting structures by @Alessan-git in #248

• Fix #250 -- qwen2vl size configuration moved by @davidkoski in #254

• update documentation by @davidkoski in #258

• Feature/generate completion info by @ronaldmannak in #259

• Decode token in generate() by @ronaldmannak in #260

• change method visibility to public. add some center crop / size related methods by @davidkoski in #263

• if the tokenizer produces an incomplete character do not consume by @davidkoski in #261

• VLM support for image and video processing with SmolVLM support by @cyrilzakka in #206

• Updated LLMEval and VLMEval with the new AsyncStream token generation. by @Alessan-git in #256

• fix various warnings -- deprecations by @davidkoski in #265

• additional changes related to async eval by @davidkoski in #266

• add some documentation on porting models by @davidkoski in #264

• Add Qwen 2.5 VL by @DePasqualeOrg in #222

• Porting doc edits by @nwh in #271

• use local pre-commit hook for swift-format by @nwh in #272

• Implement Structured Chat Messages by @ibrahimcetin in #257

• Add MLXChatExample App by @ibrahimcetin in #278

• ✨ feat: Add support for Qwen3 model by @johnmai-dev in #286

• swift-format by @davidkoski in #290

• ✨ feat: Add support for Qwen3 MoE model by @johnmai-dev in #294

• Add Qwen3 to Chat Example by @WesleyBatista in #292

• Remove unused package AsyncAlgorithms by @gtokman in #287

• Handle missing chat template and propagate other errors by @DePasqualeOrg in #288

• adding support for Granite by @smdesai in #284

• Fix UserInput init(messages:images:videos:tools:additionalContext:) broken for VLM by @adrgrondin in #298

• Refactor VLMEval to use structured messages by @ibrahimcetin in #300

• Update MLXChatExample for iOS Sandboxing by @rudrankriyam in #285

• swift-format by @davidkoski in #301

• ✨ feat: Add support for GLM-4 model by @johnmai-dev in #302

• ✨ feat: Add support for Xiaomi MiMo model by @johnmai-dev in #306

• Update READMEs of MLXLLM and MLXVLM to Use Current Registry Class Names by @rudrankriyam in #307

• chat example (command line) by @davidkoski in #277

• fix #291 -- support heterogenous quant config by @davidkoski in #293

• Fix #297 - Basic markdown support in MLXChatExample by @Kawba in #303

• Update Config for swift-transformers by @DePasqualeOrg in #322

• Fix Mutability Build Errors by @JustinMeans in #325

• adopt mlx-swift 0.25.2 by @davidkoski in #313

• consume security scoped bookmarks to read urls by @davidkoski in #314

• support for running all the models by @davidkoski in #317

New Contributors

• @finnvoor made their first contribution in #160
• @shuhongwu made their first contribution in #162
• @louen made their first contribution in #170
• @rudrankriyam made their first contribution in #169
• @JCSooHwanCho made their first contribution in #171
• @hasnat made their first contribution in #183
• @tzxdtc made their first contribution in #194
• @nvnsho made their first contribution in #187
• @ynagatomo made their first contribution in #199
• @toffaletti made their first contribution in #219
• @ibrahimcetin made their first contribution in #226
• @Alessan-git made their first contribution in #239
• @cyrilzakka made their first contribution in #206
• @nwh made their first contribution in #271
• @WesleyBatista made their first contribution in #292
• @gtokman made their first contribution in #287
• @adrgrondin made their first contribution in #298
• @Kawba made their first contribution in #303

Full Changelog: 2.21.2...2.25.4