Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Zig INferenCe Engine — Local LLM inference on AMD GPUs and Apple Silicon
| Date | Stars |
|---|---|
| 2026-07-24 | 487 |
| 2026-07-25 | 487 |
| 2026-07-28 | 487 |
| 2026-07-30 | 487 |
| 2026-07-31 | 491 |
| 2026-08-06 | 491 |
Today
— stars today
This week
+4 stars this week
This month
— stars this month
Momentum
4.0
growth rate 0.82%/day
<p align="center">
<img src="assets/zinc_trademark_new.png" alt="ZINC Logo" width="400">
</p>
# ZINC — Zig INferenCe Engine
<p align="center">
<a href="https://github.com/zolotukhin/zinc/actions/workflows/test.yml">
<img src="https://github.com/zolotukhin/zinc/actions/workflows/test.yml/badge.svg" alt="CI Status">
</a>
<a href="https://ziglang.org/download/">
<img src="https://img.shields.io/badge/Zig-0.15.2-orange.svg?logo=zig&logoColor=white" alt="Zig Version">
</a>
<a href="LICENSE">
<img src="https://img.shields.io/badge/License-MIT-blue.svg" alt="License">
</a>
<img src="https://img.shields.io/badge/platform-linux%20%7C%20macOS-lightgrey" alt="Platform">
<a href="https://zolotukhin.ai/zinc">
<img src="https://img.shields.io/badge/web-zolotukhin.ai%2Fzinc-8B5CF6" alt="Website">
</a>
<a href="https://discord.gg/QRUgWH2aGV">
<img src="https://img.shields.io/badge/Discord-Join%20ZINC-5865F2?logo=discord&logoColor=white" alt="ZINC Discord">
</a>
</p>
> Fastest measured local LLM inference for AMD GPUs. ZINC beats llama.cpp across the current five-model RDNA4 headline sweep — decode, prefill, end-to-end, and model-level overall — with no ROCm.
<p align="center">
<img src="assets/amd-rdna4-benchmark-sweep.svg" alt="AMD RDNA4 benchmark sweep chart showing ZINC ahead of llama.cpp across five models" width="860">
</p>
## AMD RDNA4: ZINC Beats llama.cpp
ZINC is the fastest measured local AI engine for AMD GPUs in our current suite.
On the Radeon AI PRO R9700, it beats llama.cpp on all five published models:
decode, prefill, end-to-end, and overall.
Same GPU. Same GGUF files. Same prompts. Same server-vs-server harness.
| Model | Decode | Prefill | Overall |
|-------|-------:|--------:|--------:|
| Qwen 3.6 35B A3B UD Q4_K_XL | **166.8** vs 108.5 tok/s (**1.54x**) | **540** vs 397 tok/s (**1.36x**) | **151%** |
| Qwen 3.5 9B Q4_K_M | **97.5** vs 85.5 tok/s (**1.14x**) | **739** vs 549 tok/s (**1.35x**) | **115%** |
| Qwen 3.6 27B Dense Q4_K_M | **32.0** vs 30.7 tok/s (**1.04x**) | **213** vs 184 tok/s (**1.16x**) | **105%** |
| Gemma 4 26B-A4B MoE Q4_K_M | **113.7** vs 102.1 tok/s (**1.11x**) | **809** vs 497 tok/s (**1.63x**) | **115%** |
| Gemma 4 31B Q4_K_M | **28.8** vs 28.5 tok/s (**1.01x**) | **249** vs 200 tok/s (**1.25x**) | **103%** |
The closest row is Gemma 4 31B decode at `1.01x`. We are still cooking: the next work is widening that margin and adding more scenarios.
## Supported Platforms
| Platform | GPU | Backend | Status |
|----------|-----|---------|--------|
| **Linux** | AMD RDNA4 (RX 9070, AI PRO R9700) | Vulkan | Primary — hand-tuned shaders |
| **Linux** | AMD RDNA3 (RX 7900 XTX, etc.) | Vulkan | Supported |
| **Linux** | Intel Arc Xe2 / Battlemage | Vulkan | Supported — validated benchmark target |
| **macOS** | Apple Silicon (M1, M2, M3, M4, M5) | Metal | Supported — native MSL shaders |
ZINC focuses on current local-inference models people are actively running:
Qwen 3.5/3.6 and Gemma 4 today, with a managed catalog that stays narrow on
purpose. Older Llama/Mistral/Gemma generations may work eventually, but broad
legacy-model coverage is not the main optimization target.
## Status vs llama.cpp
Latest checked-in benchmark artifact, same machine, same weights, same prompt:
| Platform | Compared models | Decode vs llama.cpp | Prefill vs llama.cpp | Read this as |
|----------|----------------:|--------------------:|---------------------:|--------------|
| AMD RDNA4 / Vulkan | 5 | 117% avg, 5/5 model wins | 135% avg, 5/5 model wins | Clean current sweep: every published RDNA model is ahead on decode, prefill, end-to-end, and model-level overall |
| Intel Arc / Vulkan | 5 | 103% avg, 5/5 model wins | 181% avg, 5/5 model wins | Official Linux Vulkan target; all five catalog rows validate on the public Intel node, with performance tuning still younger than RDNA4 |
| Apple Silicon / Metal | 5 | 87% avg, 1 model win | 54% avg, 1 model win | Mixed by model; GemmExcerpt of 19,607 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:b53ef7732c643ce4, topic:transformer, topic:gpt
matched fp:b53ef7732c643ce4, topic:gpu
matched fp:b53ef7732c643ce4, topic:pytorch