read more: https://www.zyphra.com/post/zuna

This is a 3B parameter model from Nanbeige with surprisingly strong benchmarks. It beats Qwen3-32B on Arena-Hard and LiveCodeBench despite being 10x smaller. (Also be warned it thinks a lot).

I wrapped it in a simple browser demo using Transformers.js + WebGPU. It downloads the q4 ONNX weights (~1.7GB) and runs fully client-side. no server required. Falls back to WASM if WebGPU isn't available.

Hugging Face model: https://huggingface.co/apple/Sharp and demo: https://huggingface.co/spaces/ronedgecomb/ml-sharp