---
title: ImpulseLeap
emoji: ⚡
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
---

# ImpulseLeap (Alexei Goncharov / RafaelUI)

ImpulseLeap is the brand of independent developer Alexei Goncharov, who builds precise, focused tools for creative professionals. Every product is made with deep care for craft — native, fast, and local-first.

On Hugging Face, the focus is machine learning that runs well on Apple Silicon: open data and tooling for the Russian-language segment and for English, with an emphasis on retrieval, SFT and RL.

## Products

All four are connected and built around one idea: optimization for Apple Silicon, from the GPU kernel up to the app.

**[Neuro](https://apps.apple.com/app/neuro-local-fine-tune/id6786251266)** — an iOS app for fine-tuning and running compact RWKV models fully on-device. Your data never leaves the phone. Models and datasets come from the catalog here on Hugging Face.

**[rwkv-metal](https://github.com/RafaelUI/rwkv-metal)** — a Python framework for RWKV-7 on Apple Silicon (`pip install rwkv-metal`): pretraining from scratch, LoRA/QLoRA, state-tuning, embeddings and reranking, built on MLX and a custom Metal WKV-7 kernel. Designed to fit 16 GB Macs.

**[rwkv-quant](https://github.com/RafaelUI/rwkv-quant)** — a quantization toolkit and Metal inference backend for RWKV-7 (`pip install rwkv-quant`). Its `.rwkvq` format makes models 2–3× smaller, with custom kernels that decode the quantized weights directly. Benchmarked against llama.cpp on a base M4 MacBook Air.

**[SwiftRWKV](https://github.com/RafaelUI/SwiftRWKV)** — a Swift package for RWKV-7 inference and training on iOS, iPadOS, macOS and visionOS. Generation, LoRA fine-tuning, reranking, embeddings and `.rwkvq` models, on the same Metal kernel. It powers Neuro.

## How they fit together

rwkv-metal trains and fine-tunes, rwkv-quant compresses and serves, SwiftRWKV brings the same stack to Swift, and Neuro puts it in your hands.

Website: https://www.impulseleap.com · GitHub: https://github.com/RafaelUI