Apple developers who’ve been eyeing large language models for their apps just got a significant boost. The team behind swift-coreml-transformers has released swift-transformers, an alpha Swift package designed to bring a transformers-like API to Apple’s ecosystem with a focus on text generation. It’s an evolution of earlier work, now sporting broader ambitions: Hub integration, arbitrary tokenizer support, and pluggable model architectures.
The release package includes more than just the library. Developers also get swift-chat, a bare-bones app showing how to wire the package into a real project, plus updated versions of exporters and transformers-to-coreml for converting PyTorch models to Core ML. Pre-converted models like Llama 2 7B and Falcon 7B are ready to drop into projects, which removes one of the biggest friction points for developers who don’t want to wrestle with conversion pipelines.
The announcement is candid about the messy parts. Conversion is “sometimes convoluted,” and the team acknowledges their tools won’t work for every new model that hits Hugging Face. Tokenization gets special attention as a genuinely hard problem — turning text into the numerical inputs models expect involves many strategies, and there’s no one-size-fits-all answer. Generation algorithms matter too: after the model predicts a probability distribution for the next token, you still need to decide which token to actually pick, and that choice shapes the quality of everything the model produces.
What’s missing is just as telling as what shipped. Not all model families are supported yet, and the post points to future work rather than pretending the problem is solved. For Swift developers who’ve watched the LLM wave mostly from the sidelines — building against cloud APIs rather than shipping models in their binaries — this is the first serious attempt to make on-device generation feel like a normal part of Apple development. The commitment to community contribution, with explicit calls for PRs and bug reports, suggests this is a foundation rather than a finished product.
