Show HN: Rust-based translator with KULYK models for English<->Ukrainian MT (github.com)

A new Rust-based English↔Ukrainian machine translation demo (Kulyk) was released, built around the llama-cpp-2 crate to run local GGUF-quantized KULYK models. Created for Rustcamp 2025 Summer, the project provides a small-footprint, self-hostable translator with a web UI (serve at http://localhost:3000) and ready Docker images for aarch64 and x86_64. Source and build scripts include cargo release via goreleaser, cross-compilation hints (cross), and a convenience task to fetch quantized models (just download_models). The implementation highlights practical, low-latency MT on consumer GPU hardware: tested on an NVIDIA GeForce RTX 3090 with 16 threads, it used ~1.2 GiB of GPU memory and averaged about 53.7 ms per sentence on a speech-uk test set. Key technical takeaways for ML practitioners are the use of GGUF quantized weights to reduce memory and improve latency, integration with llama-cpp-2 for inference in Rust, and containerized cross-platform builds for edge deployment. This makes the project a useful reference for researchers and engineers exploring privacy-preserving, offline neural MT, fast inference with quantized LLMs, and production-ready Rust toolchains.