Changelog

July 28, 2025

🛠️ Customization tools: Finetuning
- Initial release of finetuning tools for LFM2 models on ≤ 1 GPU (Google Colab notebooks) and > 1 GPU (leap-finetune)
- See more details here
📦 Model Bundling Service
- Use the Finetuning customization tools described above along with leap-bundle to develop and deploy finetuned LEAP models on the edge
- CLI-based service to create model bundles for usage within the Edge SDK - currently supports any model architecture within the LEAP model library
- See more details here
🤖 Android Edge SDK
- Features
  - Added ModelLoadingOptions and GenerationOptions for more finegrained control over generation and loading options
  - Exposed model ID via ModelRunner.modelId
  - Exposed generation statistics via stats field on MessageResponse.Complete
  - Added Model Downloader module to simplify model fetching in prototypes and development; see details here
- 🐛 Bug fixes
  - Added Proguard rules to preserve inference engine class names and prevent obfuscation issues.
  - If the generation content hits the maximum context length, finishReason field of MessageResponse.Complete will be EXCEED_CONTEXT. If the prompt exceeds the context length, LeapGenerationPromptExceedContextLengthException will be thrown.
🍎 iOS Edge SDK
- Features
  - Added ModelLoadingOptions and GenerationOptions for more finegrained control over generation and loading options
  - Exposed model ID via ModelRunner.modelId
  - Exposed generation statistics via stats field on MessageResponse.Complete
  - Added LeapModelDownloader module to simplify model fetching in prototypes and development; see details here
- 🐛 Bug fixes
  - If the generation content hits the maximum context length, finishReason field of MessageResponse.Complete will be EXCEED_CONTEXT. If the prompt exceeds the context length, LeapGenerationPromptExceedContextLengthException will be thrown.