## cactus-compute/cactus — v1.12…v1.13

_23 commits._

### Features
- **Add Whisper v3 (large-v3) support (#557)** (e65ae06)
- **Add branch override option to publish workflow (#548)** (d69695d)
- **Add cache cleanup for Hugging Face models in export_and_publish_model (#547)** (18c48f2)

### Fixes
- **Fix streaming transcribe (#576)** (130a60d)
- **fix apple i8mm detection to use runtime sysctl check (#562)** (0d44b06)
- **LFM2-VL-450M: fix garbled output (vision tower, template, kernel) (#565)** (f2385f3)
- **Update CMakeLists.txt to include gemma4 model sources and fix Dart typedefs** (97d4414)
- **Parakeet streaming fix (#551)** (aa21354)
- **Fix VLM prefill cache reuse passing all image paths for delta tokens (#545)** (dd5bd27)

### Backend
- **commit** (4cc8202)
- **Karen/needle (#574)** (f8afc46)
- **Pyannote features and optimizations (#571)** (4e54859)
- **Graph save load (#556)** (35de8df)
- **Expose min_p and repetition_penalty in completion options (#560)** (e9cc468)
- **Change default transcribe model to nvidia/parakeet-tdt-0.6b-v3 (#566)** (d4af1a1)
- **done (#567)** (6228b62)
- **Feature Cleanup (#559)** (57a6ccc)
- **Tool call prompt formatting (#558)** (c768b04)
- **refactor engine and cactus** (35495d7)
- **aggressive graph and kernel refactor** (aa23d7e)
- **Update blog URL in README.md (#544)** (f03da54)
- **Use raw state dict loading for Gemma 3n conversion (#549)** (6273c48)
- **Improve model publishing error handling in workflow (#546)** (8647c25)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/cactus-compute/cactus?utm_source=github-action)._