v0.5.0

• Add support for downloading and inferencing with Gemma 3 4B 3-bit quantization model
• Simplify LLM static accessors by removing duplicate most duplicate accessors (e.g., gemma2_2B and gemma2_9B are combined into gemma2)

v0.4.3

• Add SHLLM.cacheLimit, SHLLM.memoryLimit, and SHLLM.recommendedMaxWorkingSetSize
• Remove processing: UserInput.Processing argument from Gemma 3 initializers

v0.4.2

• Update mlx-swift-examples
• Add processing: UserInput.Processing? argument to LLM.init()

v0.4.1

• Re-export LanguageModel, LLMModel, and VLMModel

v0.4.0

• Add Gemma 3 1B (text-only) and 4B, 12B, and 27B (vision models)
• Add support for new tool-calling mechanism
• Fix crash when models exceed output token count

v0.3.0

• Update to newest release of mlx-swift-examples
• Use new, simplified MLX method of loading and initializing models
• Improve algorithm for limiting maximum number of input tokens

v0.2.0

• Add support for Qwen3 models.

v0.1.0

• Convert LLM into an AsyncSequence instead of an actor.

let input: UserInput = .init(messages: [
  ["role": "system", "content": "You are a helpful assistant."],
  ["role": "user", "content": "What is the meaning of life?"],
])

guard let llm = try qwen2_5__7B(input) else { return }
for try await reply in llm {
  print(reply)
}

v0.0.6

• Add LLM.clearCache()

v0.0.5

• Export Message and UserInput types.